Pith. sign in

REVIEW 5 major objections 8 minor 107 references

Real-Time Anomaly Detection in Video Streams

T0 review · 5 major / 8 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This thesis claims that fusing YOLO object detection with a VGG19-GRU temporal network can spot fights, gunshots, and fires in video streams with up to 90 percent precision and recall, while also providing visual explanations for each…

desk verdict A thorough engineering thesis with honest limitations, but the headline accuracy numbers don't survive contact with its own confusion matrices. read the letter →

arxiv 2411.19731 v1 pith:72CM5MLD submitted 2024-11-29 cs.CV cs.LG

Fabien Poirier This is my paper · ORCID
classification cs.CVcs.LG
keywords anomalydetectionvideostreamsreal-timeYOLOVGG19gatedrecurrentunitobjectexplainability
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This thesis aims to show that dangerous events in video streams—fights, gunshots, and fires—can be detected reliably by combining two complementary analyses: spatial object detection and temporal sequence analysis. The author proposes an architecture that pairs YOLO, which locates people, firearms, and flames in each frame, with a convolutional recurrent network (VGG19 followed by a GRU) that reads 20-frame sequences to decide whether an anomaly is happening. The system is trained and evaluated on two proprietary datasets built from smartphone and web videos, and its reported accuracy reaches up to 90 percent precision and recall in certain classes and configurations. The motivation is practical: automated assistance for surveillance operators who currently monitor many cameras at once, where speed and interpretability matter.

What carries the argument

The load-bearing mechanism is a Convolutional Gated Recurrent Unit (CGRU): VGG19, applied through a time-distributed wrapper to each frame of a 20-image sequence, extracts spatial features, and a GRU with a multi-layer perceptron classifier reads the resulting feature sequence to label it as normal or one of the anomaly types. This temporal core is combined, in parallel mode, with YOLOv7 object detection and hand-written correction rules—a detected flame overrides the predicted class to fire, and a firearm whose bounding box has nonzero intersection over union with a person overrides to gunshot. The correction rules are the explicit coupling that lets object presence influence the temporal decision.

What would settle it

Take the trained system as described and run it on a held-out set of genuine surveillance-camera recordings containing fights, gunshots, and fires that were not part of the author's datasets; if precision and recall fall far below the reported 90 percent, the claim of real-world anomaly detection is refuted. A simpler check is to remove the hand-written correction rules and observe whether performance on fire and gunshot classes collapses, which would show the temporal model alone does not carry the result.

Watch

Extended reading notes

Core claim

The central claim is that combining YOLOv7 object detection with a temporal model built from VGG19 and a gated recurrent unit (GRU) produces a video anomaly detection system that can identify fights, gunshots, and fires with high accuracy on its own curated videos, in both whole-video and per-sequence evaluations. The author argues that the fusion of spatial information (the presence and position of key objects) with temporal dynamics (how those objects move) is what makes the detection work, and that a binary normal/abnormal model generalizes better than a multi-class model when the goal is simply to raise an alert. The system also provides per-frame saliency and activation maps, adapted to the time-distributed convolution, so an operator can see which regions drove the decision.

Load-bearing premise

The entire evaluation rests on the assumption that the author's two proprietary datasets—images and videos collected from smartphones and the web—are representative enough of real surveillance conditions for fights, gunshots, and fires that a model trained on them will behave the same way on genuine surveillance footage.

Editorial extensions

If this is right

  • A binary normal/abnormal model outperforms the multi-class model, so for pure alerting the recommended configuration is binary classification combined with YOLO.
  • Serial mode, where YOLO preprocesses frames before the CGRU, is more reliable, while parallel mode is faster; neither processes video at full frame rate, with average detection times below about 1.5 seconds.
  • Per-sequence evaluation, which corresponds to continuous-stream monitoring, gives lower accuracy than whole-video evaluation, so the system is better at classifying completed clips than at flagging events mid-stream.
  • Specialized single-anomaly models beat the multi-class model for a given anomaly type, while the multi-class model remains viable when the precise type must be identified.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the correction rules were replaced with a learned fusion module, the system might adapt more gracefully to contexts where a weapon is present but not dangerous, such as airport security—an example the author cites as a known failure mode.
  • The reported per-sequence accuracy suggests the 20-frame window and dynamic-step generator may be a limiting factor for true streaming detection; testing shorter or overlapping windows on the same data would isolate whether window length or model capacity is the bottleneck.
  • Because the datasets are proprietary and collected from online videos, the results are not directly comparable to public benchmarks; a natural next step would be to retrain the same architecture on a public anomaly dataset and compare per-class numbers.
  • The explainability module, which highlights low-activation gun regions, could be turned into a weakly supervised localization signal to eventually output bounding boxes around the anomaly, which the author lists as a missing feature.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 8 minor

Summary. This manuscript, a 2023 doctoral thesis posted to arXiv, presents a video anomaly-detection system that couples YOLO object detection (spatial analysis) with a VGG19-GRU temporal classifier ('CGRU'), usable in serial or parallel modes, extended with hand-crafted correction rules (flame implies fire; firearm overlapping a person implies gunshot) and with explainability outputs (saliency and activation maps adapted to video, a contour method, and a 'badBox' evaluation metric). Two proprietary datasets were assembled, one of images (firearms, flames, humans) and one of videos (fights, gunshots, fires versus normal). On its validation videos the system is reported to reach up to 90% precision and recall (e.g., Table 4.43, fire/normal sequence detection), and the Conclusion summarizes the system as providing 'very good results, reaching up to 90% precision and recall in certain cases.' The title claims real-time detection in video streams.

Significance. If the reported results were robust, this would be a useful engineering contribution: an interpretable surveillance-assistance pipeline with two deployment modes, documented per-decision latencies, and a notably candid limitations section (Conclusion, Section 10, admits the absence of surveillance-camera footage and the non-real-time operation). The experimental breadth is a genuine strength: C3D, ConvLSTM, and CNN+RNN variants are compared across several backbones (Tables 4.27-4.31), model selection is documented in detail, the extension of saliency and activation maps to time-distributed video models is concretely described, and the badBox metric is an original evaluation element. However, the central claims are not established at the level asserted: all reported numbers come from the same split used for model selection, several reported metrics cannot be reproduced from the published confusion matrices, the real-time claim contradicts the author's own conclusion, and there is no evaluation on public benchmarks or any statistical analysis.

major comments (5)
  1. [Ch. 3, Section 4.3, Tables 3.1-3.2; Ch. 4, Section 6; Conclusion, Section 10] All reported quantitative results are computed on the split called 'Validation' in Tables 3.1 and 3.2, and the same split is used in Ch. 4, Section 6.1 to select the image size (Table 4.1), the sequence length (Table 4.2), the VGG19 retraining layer (Tables 4.28-4.29), the recurrent cell (Table 4.31), the classifier depth (Table 4.32), and the GRU size and dropout (Figures 4.6-4.9). No separate test split is defined anywhere in the dataset description, and the only data that are truly held out are the ten unedited videos discussed qualitatively in Section 4.6.2. Conclusion, Section 10 confirms that 'performance metrics were thus based solely on videos from our test set,' which is the same validation split. Because this split was used for hyperparameter selection, the reported precision and recall figures (including the 'up to 90%' headline) are selection-fitted estimates on the validation distribution, not estimates of generalization to unseen data; the paper must either evaluate on a genuinely held-out test set or scale its claims down accordingly.
  2. [Tables 4.33, 4.35, 4.40, 4.43 (and 4.34)] The reported metrics cannot be reconciled with the published confusion matrices. The matrices are row-normalized (rows sum to 100%), so the diagonal entry of a class row is that class's recall. In Table 4.33 (fight, sequence detection) the reported recall is 60.3% while the matrix diagonal is 78.3%, and the reported tuple (precision 93.6%, recall 60.3%, accuracy 63.1%) is impossible with that matrix under any class prior: matching the reported accuracy forces a normal:fight prior of about 5:1, which yields a precision near 27%, not 93.6%. Table 4.35 (fire, sequence detection) reports recall 96.2% against a diagonal of 60.5%; Table 4.43 (fire/normal) reports recall 90.03% (sequence) and 87.87% (video) against diagonals of 83.67% and 93.33%; and Table 4.40 (normal/abnormal, sequence) reports precision 94.82% when the accuracy-consistent prior gives about 55%. In Table 4.34 the reported recall (84.8%) agrees with the matrix, but the reported precision (35.1%) requires an undocumented evaluation-set prior of roughly 20 normal sequences per gunshot sequence. These are not rounding effects, and since these tables are the only quantitative evidence for the central claim, the metrics, the normalization convention, and the class priors must be restated and verified.
  3. [Title; Ch. 1, Section 5; Ch. 4, Section 5; Conclusion, Section 9] The title's claim of real-time detection is contradicted by the manuscript itself. Ch. 1, Section 5 sets the goal as near real-time, Ch. 4, Section 5 states that the two parallel models are 'not yet capable of processing videos in real time,' and Conclusion, Section 9 states that 'the anomaly is detected a posteriori. It therefore does not operate in real-time.' The timing tables (4.25-4.26) show per-decision times of 0.5-1.1 s in parallel mode and 1.0-1.5 s in serial mode, with total processing times of the same order as (or longer than) the video duration, and Table 4.44 shows the same pattern on the ten held-out videos. The title and abstract should be changed to a defensible statement such as 'near-real-time anomaly detection,' with the measured latency reported explicitly, or the authors must demonstrate processing at a rate faster than the input frame rate.
  4. [Tables 4.6, 4.20, 4.23-4.24] Tables 4.23 and 4.24 are verbatim duplicates, including the erroneous 'RCNN + YOLO' caption, so the 'reducing false positives' experiment is reported twice as two different tables. Moreover, Table 4.6 reports exactly identical metrics for CGRU alone and for CGRU fed with YOLO bounding-box overlays on the gunshot/normal task (accuracy 91.89%, precision 35.14%, recall 84.86%, F1 49.70%), and Table 4.20 repeats those identical values, with the text conceding that 'displaying the bounding boxes around objects has no impact on the performance of our model.' Since the combination of spatial and temporal analysis is one of the two stated contributions, the serial-mode fusion is shown to be ineffective for the gunshot class, and the gains in parallel mode come from the hand-crafted correction rules rather than from learned fusion; the contribution claim should be re-scoped to the configurations that actually change the results, and the duplicate tables should be removed.
  5. [Ch. 3, Sections 5 and 7; Ch. 4, Sections 6-7; Conclusion, Section 10] The evaluation is a single-run, single-domain assessment: there are no error bars, no repeated training runs, no significance tests, no public benchmark results (UCF Crime, ShanghaiTech, Hockey Fight, and AIRTLab are surveyed in Chapter 1 but never used), and the datasets are proprietary, so the results cannot be checked by other researchers. With only 64 gunshot and 61 fire videos in the validation split (Table 3.1), the 'up to 90%' claim is not robustly supported. In addition, the parallel-mode improvement rests on deterministic rules (flame present implies fire; firearm with IoU greater than 0 with a person implies gunshot; Ch. 3, Sections 5 and 7) whose thresholds (confidence 55%, IoU > 0) are asserted without sensitivity analysis and which the author concedes (Conclusion, Section 10) misclassify legitimate situations such as armed personnel in airports. An ablation and threshold study is needed, and the stated scope should reflect this limitation.
minor comments (8)
  1. [Throughout] Typos and terminology: 'Convolutional Neuronal Network,' 'Recurrent Neronal Network,' 'You Only Looks Once,' 'YOLO+CRU' (Table 4.6), 'RCNN + YOLO' (Tables 4.23-4.24), 'A V AD' (Ch. 1, Section 3), and 'Segment Anything Mode' (Conclusion, Section 11) should be corrected.
  2. [Ch. 2, Ch. 3, Ch. 4] Cross-references are broken ('Section ??' in Ch. 4, Section 2; 'see Section 6' in Ch. 3, Section 4; 'page 23' in Ch. 3, Section 4.3), and figure numbering is duplicated (Figure 2.1 appears in both Chapter 2 and Chapter 3; Figure 4.2 appears twice), which will confuse readers.
  3. [Tables 4.33-4.43] The 'Elements' column is never defined (number of videos versus number of sequences), and the confusion matrices do not state whether percentages are normalized by row or by column; both must be specified for the metrics to be verifiable, and the number of decimal places should be made consistent.
  4. [Table 3.1] The dataset table lists Fight, Gunshot, and Fire but omits the number of Normal videos, although Normal is one of the two classes in most of the reported binary experiments; the Normal counts should be added.
  5. [Table 4.3] The 'X' entries for Optical Flow and Mask Optical Flow are unexplained; the caption should note that the model could not be trained with these preprocessing options.
  6. [Introduction; Ch. 2, Section 4] The claimed novelty ('novel approach combining temporal and spatial analysis'; an 'original method' for explainability) should be positioned against the already-surveyed combined approaches (Doshi and Yilmaz 2020; Mustafa 2022) and against the author's own prior publications (Poirier et al. 2022, 2023a, 2023b), so that the incremental contribution is clear.
  7. [Conclusion, Section 9] The summary sentence 'reaching up to 90% precision and recall in certain cases' should name the specific configuration (Table 4.43, fire/normal, sequence detection) and should carry the caveat that the numbers come from the validation split; in its current form it overstates what the data show.
  8. [Conclusion] The Conclusion chapter numbers its sections 9, 10, and 11 while the table of contents lists only an unnumbered 'Conclusion'; the headings should be renumbered consistently.

Circularity Check

2 steps flagged · score 6.0 of 10

Reported precision/recall up to 90% are computed on the validation split used for model selection, and parallel-mode anomaly labels are defined by the YOLO object dictionary.

  1. fitted input called prediction [Table 3.1; Section 4.6.1; Conclusion Section 10]
    "Table 3.1: Video Distribution in Our Dataset Classes Train Validation Total; Fight 587 391 978; Gunshot 247 64 311; Fire 237 61 298 [...] Finally, we used the results obtained along with the explainability techniques mentioned earlier (section 6) to determine the number of neurons required for our GRU and the dropout rate for our dropout layers. [...] Our performance metrics were thus based solely on videos from our test set."

    Table 3.1 defines only Train and Validation splits; there is no held-out test split. Section 4.6.1 selects the sequence length, GRU neuron count, dropout rate, classifier depth, and VGG retraining layer by optimizing results obtained on the Validation split. Section 4.6.2 then reports the final precision/recall tables on the same Validation split, and the Conclusion calls it 'our test set'. The 'up to 90% precision and recall' claim is therefore a report of the selected model's fit on the model-selection split, not a prediction on unseen data; the only truly external evaluation (ten videos, Section 4.6.2) is qualitative and reports no precision/recall.

  2. self definitional [Chapter 3, Section 5 (Correction Component); Section 7 (Operation)]
    "we have created a dictionary where each object type learned by YOLO is linked to an anomaly: flames with the 'fire' anomaly and firearms with the 'gunshot' anomaly. [...] If no anomaly is detected but a key object known by the model is present in the sequence, then the class is defined as the anomaly associated with that key object."

    The parallel-mode 'correction' rules define the output class directly from YOLO's object classes: a detected flame is relabeled 'fire', and a detected firearm with IoU>0 against a person is relabeled 'gunshot'. Thus the anomaly label is, by construction, the object detector's output passed through a hand-made dictionary; the system cannot fail at 'anomaly detection' for these classes except by YOLO failing at object detection. This makes the parallel-mode detection result equivalent to the YOLO object-detection result plus a fixed rule, and the rule itself was chosen to match the three anomaly classes in the author's proprietary dataset.

full rationale

The central numerical claim (up to 90% precision/recall, from Tables 4.33-4.43) is not a held-out prediction: the dataset is partitioned into Train and Validation only (Table 3.1), and Section 4.6.1 uses the Validation split to choose the sequence length, GRU neuron count, dropout, classifier depth, and VGG retraining layer; the final metrics in Section 4.6.2 are then reported on the same split, which the Conclusion relabels as a 'test set'. This is fitted-input-called-prediction circularity. Additionally, the parallel-mode correction rules define the 'fire' and 'gunshot' anomaly classes by direct mapping from YOLO's flame and firearm detections (plus IoU with persons), so for those modes the anomaly label is the object detector output relabeled by a hand-made dictionary. Self-citations (Poirier et al. 2022, 2023a, 2023b) are descriptive of contributions and are not load-bearing for the accuracy claim. No uniqueness theorem or ansatz-smuggling is involved. Score 6 reflects that the central accuracy claim reduces to an in-sample fit plus definitional rules, while the temporal CGRU architecture itself is an independent construction.

Assumptions & free parameters 9 free parameters · 4 assumptions · 0 invented entities

All central results depend on hyperparameters tuned on the proprietary validation set and on domain assumptions about the representativeness of the data and the validity of the hand-coded correction rules. No public benchmark, code, or data release is provided, so the fitted values and assumptions cannot be checked independently.

free parameters (9)
  • Sequence length = 20 frames
    Selected from {15, 20, 30} as giving best validation F1 (Table 4.2).
  • Input image size = 112x112
    Selected from {80, 112, 140} on validation (Table 4.1).
  • YOLO confidence threshold = 55%
    Hand-set for all object detection evaluations (Tables 4.45-4.49).
  • NMS overlap threshold = 70%
    Hand-set for evaluation (Tables 4.45-4.46).
  • VGG19 retrained layer position = Layer 11
    Chosen via grid of retrained layers (Table 4.28).
  • GRU hidden size and dropout = Not clearly specified in final model
    Selected using activation map inspection (Figures 4.6-4.9); final values not stated in the text.
  • Classifier fully connected layers = 3 + output
    Chosen from {2, 3, 4, 5} + output (Table 4.32).
  • Gunshot decision IoU threshold = greater than 0
    Firearm-person overlap greater than 0 triggers gunshot; this rule is hand-set (Chapter 3, sections 5 and 7).
  • Data augmentation multiplier = 3x
    Videos multiplied by three via mirror, zoom, and brightness changes (Chapter 4, section 2).
assumptions (4)
  • standard math Backpropagation and gradient-based saliency/activation maps correctly attribute model decisions.
    Used throughout Chapter 4 for model selection and explainability without proof.
  • domain assumption The proprietary image and video datasets are representative of real-world fights, gunshots, and fires in surveillance conditions.
    All performance numbers rest on this; the thesis notes it could not test on real surveillance camera footage (Conclusion section 10).
  • domain assumption YOLO's object detections (person, firearm, flame) are accurate enough to serve as reliable correction signals.
    The correction component changes CGRU labels based solely on YOLO detections (Chapter 3, section 5).
  • ad hoc to paper A firearm overlapping a person (IoU greater than 0) indicates a gunshot anomaly.
    Hand-crafted rule in the parallel mode correction component (Chapter 3, section 7); the thesis itself notes this is wrong in airports.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Real-Time Anomaly Detection in Video Streams." pith.science (2026). https://pith.science/paper/72CM5MLD

@misc{pith2026241119731,
  author       = {Pith},
  title        = {Pith review of: Real-Time Anomaly Detection in Video Streams},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/72CM5MLD}},
  note         = {Machine review of arXiv:2411.19731}
}
read the original abstract

This thesis is part of a CIFRE agreement between the company Othello and the LIASD laboratory. The objective is to develop an artificial intelligence system that can detect real-time dangers in a video stream. To achieve this, a novel approach combining temporal and spatial analysis has been proposed. Several avenues have been explored to improve anomaly detection by integrating object detection, human pose detection, and motion analysis. For result interpretability, techniques commonly used for image analysis, such as activation and saliency maps, have been extended to videos, and an original method has been proposed. The proposed architecture performs binary or multiclass classification depending on whether an alert or the cause needs to be identified. Numerous neural networkmodels have been tested, and three of them have been selected. You Only Looks Once (YOLO) has been used for spatial analysis, a Convolutional Recurrent Neuronal Network (CRNN) composed of VGG19 and a Gated Recurrent Unit (GRU) for temporal analysis, and a multi-layer perceptron for classification. These models handle different types of data and can be combined in parallel or in series. Although the parallel mode is faster, the serial mode is generally more reliable. For training these models, supervised learning was chosen, and two proprietary datasets were created. The first dataset focuses on objects that may play a potential role in anomalies, while the second consists of videos containing anomalies or non-anomalies. This approach allows for the processing of both continuous video streams and finite videos, providing greater flexibility in detection.

Figures

Figures reproduced from arXiv: 2411.19731 by the authors.

Figure 4.1
Figure 4.1. Excerpt from the UCSD dataset UCF Crime is currently the benchmark dataset for anomaly detection in videos. It was created by Sultani, C. Chen, and Shah (2018) at the University of Florida and contains 1,900 untrimmed videos representing realistic anomalies across 13 categories: abuse, arrests, arson, assault, road accidents, burglary, explosions, fighting, robbery, shootings, shoplifting, stealing, and vandalism, a… view at source ↗
Figure 1.1
Figure 1.1. Recurrent Architectures (Toharudin et al., [PITH_FULL_IMAGE:figures/full_fig_p030_1_1.png] view at source ↗
Figure 1.2
Figure 1.2. VGG19 Architecture (Lagunas and Garces, 2018) Other architectures are also used in image processing. For example, ResNet (He et al., 2015) is a deep network based on the idea of “shortcuts.“ The addition of resid￾ual connections helps prevent performance degradation due to additional layers. Inception (Szegedy, W. Liu, et al., 2014) uses inception blocks to extract features at different spatial scales from the same … view at source ↗
Figures from the paper (59 more)
Figure 1.3
Figure 1.3. Figure 1.3: Operation of convolution filters, Ji et al. ( [PITH_FULL_IMAGE:figures/full_fig_p033_1_3.png]
Figure 1.4
Figure 1.4. Figure 1.4: Architecture of an autoencoder, Sublime ( [PITH_FULL_IMAGE:figures/full_fig_p034_1_4.png]
Figure 1.5
Figure 1.5. Figure 1.5: Image autoencoder, Chollet et al. (2015) 1.4 Temporal Convolutional Network (TCN) Temporal Convolutional Networks (TCN) are a family of convolutional neural network models designed for time-series modeling ( [PITH_FULL_IMAGE:figures/full_fig_p035_1_5.png]
Figure 1.6
Figure 1.6. Figure 1.6: Architecture of TCN (Lea et al., 2016) Dilated filters play a crucial role in the ability of TCNs to capture temporal pat￾terns at different scales. Dilated filters are convolutions where the indices of the values to be considered are spaced by a specific number of s…
Figure 1.7
Figure 1.7. Figure 1.7: Architecture of the TCN-based autoencoder (Lea et al., [PITH_FULL_IMAGE:figures/full_fig_p036_1_7.png]
Figure 1.8
Figure 1.8. Figure 1.8: Transformer architecture, Vaswani et al. ( [PITH_FULL_IMAGE:figures/full_fig_p038_1_8.png]
Figure 1.9
Figure 1.9. Figure 1.9: Vision Transformer, P. Wang (2020) and Hassani et al. (2021) Fabien Poirier Paris 8 University [PITH_FULL_IMAGE:figures/full_fig_p040_1_9.png]
Figure 1.10
Figure 1.10. Figure 1.10: Vision Transformer, B. Wu et al. (2020) [PITH_FULL_IMAGE:figures/full_fig_p041_1_10.png]
Figure 1.11
Figure 1.11. Figure 1.11: Attention map generated by DINO, Caron et al. ( [PITH_FULL_IMAGE:figures/full_fig_p041_1_11.png]
Figure 1.12
Figure 1.12. Figure 1.12: Vanilla GAN Architecture, Cai et al. (2020) S. Zhu, C. Chen, and Sultani (2020) provide a comprehensive discussion on super￾vised and unsupervised deep learning methods for anomaly detection in surveil￾lance videos. Additionally, Ghosh (2018) explore action recognit…
Figure 2.1
Figure 2.1. Figure 2.1: Boost Cascade (Viola and Jones, 2001) Later, other technologies emerged, such as Histogram of Oriented Gradients (HOG) and Deformable Parts Model (DPM), all relying on manually selected feature extraction techniques, including edges, corners, and gradients in images,…
Figure 2.2
Figure 2.2. Figure 2.2: RCNN operation, Girshick et al. (2013) First, the input image is divided into 2000 sub-images called “regions of inter￾est“ using the method known as “Selective Search” (Uijlings et al., 2013). This method divides the input image into several parts, and each neighbor…
Figure 2.3
Figure 2.3. Figure 2.3: How YOLO works, Redmon, Divvala, et al. ( [PITH_FULL_IMAGE:figures/full_fig_p045_2_3.png]
Figure 2.4
Figure 2.4. Figure 2.4: Non-Maximum Suppression (NMS), Jain and Nandy ( [PITH_FULL_IMAGE:figures/full_fig_p046_2_4.png]
Figure 2.5
Figure 2.5. Figure 2.5: DenseNet architecture, Huang, Z. Liu, and Weinberger ( [PITH_FULL_IMAGE:figures/full_fig_p047_2_5.png]
Figure 2.6
Figure 2.6. Figure 2.6: DenseNet vs CSPDenseNet, C.-Y. Wang, H.-Y. M. Liao, et al. ( [PITH_FULL_IMAGE:figures/full_fig_p048_2_6.png]
Figure 2.7
Figure 2.7. Figure 2.7: YOLOV4 data augmentation, Bochkovskiy, C.-Y. Wang, and H.-Y. M. [PITH_FULL_IMAGE:figures/full_fig_p048_2_7.png]
Figure 2.8
Figure 2.8. Figure 2.8: YOLOF architecture, Q. Chen et al. (2021) [PITH_FULL_IMAGE:figures/full_fig_p049_2_8.png]
Figure 2.9
Figure 2.9. Figure 2.9: YOLOP architecture, D. Wu et al. (2021) • YOLOR (You Only Learn One Representation) uses both explicit knowledge (from its learning) and implicit knowledge. Thanks to the implicit knowl￾edge acquired, this model performs very well for multitask learning, and can ther…
Figure 2.10
Figure 2.10. Figure 2.10: YOLOR architecture, C.-Y. Wang, Yeh, and H. Liao ( [PITH_FULL_IMAGE:figures/full_fig_p050_2_10.png]
Figure 2.11
Figure 2.11. Figure 2.11: YOLOS architecture (Fang et al., 2021) • (Ge et al., 2021) introduced YOLOX, a variant of YOLOV3 that does not use any anchor boxes, resulting in faster and more accurate performance. 2.2.6 YOLOV5 A few months after the release of YOLOV4, version 5 was released1 ; h…
Figure 2.12
Figure 2.12. Figure 2.12: Architectures proposed by YOLOV5 2.2.7 YOLOV6 Recently, the Chinese company Meituan released MT-YOLOV61 , an algorithm that is not officially part of the YOLO series but is heavily inspired by the original YOLO. It has no relation to version 5, but it surpasses it s…
Figure 2.13
Figure 2.13. Figure 2.13: Performance of YOLOV7, C.-Y. Wang, Bochkovskiy, and H.-Y. M. Liao (2022) [PITH_FULL_IMAGE:figures/full_fig_p053_2_13.png]
Figure 2.14
Figure 2.14. Figure 2.14: Applications of YOLOV7, C.-Y. Wang, Bochkovskiy, and H.-Y. M. Liao (2022) More details about the functioning of these models can be found in Sharma (2018c), Anomaly Detection in Videos 2023 [PITH_FULL_IMAGE:figures/full_fig_p053_2_14.png]
Figure 2.15
Figure 2.15. Figure 2.15: Evolution of object detection (Zou et al., [PITH_FULL_IMAGE:figures/full_fig_p054_2_15.png]
Figure 2.1
Figure 2.1. Figure 2.1: Overall Architecture In the following sections, we will explain each component of the figure above (the diamonds), before summarizing the overall functioning of the system. 3 Spatial Analysis Component The first element of our system performs a spatial analysis of ea…
Figure 4.1
Figure 4.1. Figure 4.1: Architecture of the temporal analysis module (CGRU) [PITH_FULL_IMAGE:figures/full_fig_p065_4_1.png]
Figure 4.2
Figure 4.2. Figure 4.2: Data Distribution 5 Correction Component The fourth component of our system is used to combine, when operating in parallel mode, the object detection results with those of the temporal analysis. Its role is Anomaly Detection in Videos 2023 [PITH_FULL_IMAGE:figures/f…
Figure 6.1
Figure 6.1. Figure 6.1: 2D Convolution / 3D Convolution / Convolution embedded in a Time [PITH_FULL_IMAGE:figures/full_fig_p069_6_1.png]
Figure 7.1
Figure 7.1. Figure 7.1: Overall Operation Fabien Poirier Paris 8 University [PITH_FULL_IMAGE:figures/full_fig_p072_7_1.png]
Figure 2.1
Figure 2.1. Figure 2.1: Functioning of the tested generators (Ferlet, [PITH_FULL_IMAGE:figures/full_fig_p077_2_1.png]
Figure 2.2
Figure 2.2. Figure 2.2: Data augmentation [PITH_FULL_IMAGE:figures/full_fig_p079_2_2.png]
Figure 2.3
Figure 2.3. Figure 2.3: Preprocessing applied to an image representing a gunshot [PITH_FULL_IMAGE:figures/full_fig_p079_2_3.png]
Figure 2.4
Figure 2.4. Figure 2.4: Images representing a fight Fabien Poirier Paris 8 University [PITH_FULL_IMAGE:figures/full_fig_p080_2_4.png]
Figure 2.5
Figure 2.5. Figure 2.5: Attention map for a gunshot video class After testing DINO on our various datasets, we found that this method requires sig￾nificant computational time, making it unsuitable for our near-real-time problem. However, the results obtained on gunshot videos were encouragi…
Figure 2.6
Figure 2.6. Figure 2.6: Overfitting for the Fight class with DINO preprocessing [PITH_FULL_IMAGE:figures/full_fig_p082_2_6.png]
Figure 3.1
Figure 3.1. Figure 3.1: Mask created with YOLO For any image where no object is detected, we had two options: either keep the original image or replace it with a black image. Since this preprocessing is crucial for the reliability of our anomaly detection model, we decided to test both poss…
Figure 3.2
Figure 3.2. Figure 3.2: Pose estimation by YOLOV7 for a fight, without background [PITH_FULL_IMAGE:figures/full_fig_p085_3_2.png]
Figure 3.3
Figure 3.3. Figure 3.3: Pose estimation by YOLOV7 for a fight, with background Fabien Poirier Paris 8 University [PITH_FULL_IMAGE:figures/full_fig_p086_3_3.png]
Figure 4.1
Figure 4.1. Figure 4.1: Result of YOLO V4 trained on the classes gun + person [PITH_FULL_IMAGE:figures/full_fig_p090_4_1.png]
Figure 4.2
Figure 4.2. Figure 4.2: Activation maps for a video from the gunshot class extracted from the [PITH_FULL_IMAGE:figures/full_fig_p094_4_2.png]
Figure 4.3
Figure 4.3. Figure 4.3: Example visualization for the classes: gunshot and fight [PITH_FULL_IMAGE:figures/full_fig_p094_4_3.png]
Figure 4.4
Figure 4.4. Figure 4.4: Activation map and contour visualization on an image from the gunshot [PITH_FULL_IMAGE:figures/full_fig_p095_4_4.png]
Figure 4.5
Figure 4.5. Figure 4.5: Activation map and contour visualization on an image from the fight [PITH_FULL_IMAGE:figures/full_fig_p095_4_5.png]
Figure 4.6
Figure 4.6. Figure 4.6: Activation maps showing the variation of dropout using 64 GRU neu [PITH_FULL_IMAGE:figures/full_fig_p096_4_6.png]
Figure 4.7
Figure 4.7. Figure 4.7: Activation maps for a variation in the number of GRU neurons using [PITH_FULL_IMAGE:figures/full_fig_p096_4_7.png]
Figure 4.8
Figure 4.8. Figure 4.8: Activation maps for a variation in the number of classifier layers [PITH_FULL_IMAGE:figures/full_fig_p096_4_8.png]
Figure 4.9
Figure 4.9. Figure 4.9: Activation maps of our model with 1024 GRU neurons and 0% dropout [PITH_FULL_IMAGE:figures/full_fig_p097_4_9.png]
Figure 4.10
Figure 4.10. Figure 4.10: Visualization of a normal video from the test set representing a rental [PITH_FULL_IMAGE:figures/full_fig_p098_4_10.png]
Figure 4.11
Figure 4.11. Figure 4.11: Saliency map for a normal video. Fabien Poirier Paris 8 University [PITH_FULL_IMAGE:figures/full_fig_p098_4_11.png]
Figure 4.12
Figure 4.12. Figure 4.12: 4 convolution filters for layers 1 and 2. [PITH_FULL_IMAGE:figures/full_fig_p099_4_12.png]
Figure 6.1
Figure 6.1. Figure 6.1: Fight detection, Source: Le Parisien [PITH_FULL_IMAGE:figures/full_fig_p117_6_1.png]
Figure 6.2
Figure 6.2. Figure 6.2: Fire detection, Source: Le Parisien [PITH_FULL_IMAGE:figures/full_fig_p117_6_2.png]
Figure 6.3
Figure 6.3. Figure 6.3: Gunshot detection [PITH_FULL_IMAGE:figures/full_fig_p117_6_3.png]
Figure 7.1
Figure 7.1. Figure 7.1: Learning Curve for YOLOV4: 2 Classes Fabien Poirier Paris 8 University [PITH_FULL_IMAGE:figures/full_fig_p120_7_1.png]
Figure 7.2
Figure 7.2. Figure 7.2: Learning Curve for YOLOV4: 1 Class 7.2 Results After achieving satisfactory results with the YOLOV4 model trained on a single class, we evaluated the model’s performance under anomaly detection conditions. Our initial approach was to define an anomaly as a detected f…
Figure 7.3
Figure 7.3. Figure 7.3: Gun Detection via YOLOV4 Fabien Poirier Paris 8 University [PITH_FULL_IMAGE:figures/full_fig_p122_7_3.png]
Figure 7.4
Figure 7.4. Figure 7.4: Flame Detection via YOLOV4 8 Conclusion Our tests allowed us to define a set of models with the following common features: the choice of YOLOV7 (even though some of our tests were conducted before this version was released and used V4); the use of GRU and VGG19 convo…
Figure 10.1
Figure 10.1. Figure 10.1: Accident Detection Despite good performance from our system and acceptable processing times that can still be improved, we have not yet been able to target the area where Fabien Poirier Paris 8 University [PITH_FULL_IMAGE:figures/full_fig_p128_10_1.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

107 extracted references · 63 canonical work pages

  1. [1]

    Variational autoencoder based anomaly detection using reconstruction probability

    J. An and S. Cho. “Variational autoencoder based anomaly detection using reconstruction probability”. In: Special Lecture on IE 2.1 (2015), pp. 1–18 (cit. on p. 34)

    2015

  2. [2]

    3D-CNN-Based Fused Feature Maps with LSTM Applied to Action Recognition

    S. Arif, J. Wang, T. U. Hassan, and Z. Fei. “3D-CNN-Based Fused Feature Maps with LSTM Applied to Action Recognition”. In: Future Internet 11 (2019), p. 42 (cit. on p. 24)

    2019

  3. [3]

    ViViT: A Video Vision Transformer

    A. Arnab, M. Dehghani, G. Heigold, C. Sun, M. Lucic, and C. Schmid. “ViViT: A Video Vision Transformer”. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV) (2021), pp. 6816–6826 (cit. on p. 39)

    2021

  4. [4]

    N. Barla. A Comprehensive Guide to Human Pose Estimation. 2022. url: https://www.v7labs.com/blog/human-pose-estimation-guide (cit. on p. 85)

    2022

  5. [5]

    Understanding the role of individual units in a deep neural network

    D. Bau, J.-Y. Zhu, H. Strobelt, `A. Lapedriza, B. Zhou, and A. Torralba. “Understanding the role of individual units in a deep neural network”. In: Proceedings of the National Academy of Sciences 117 (2020), pp. 30071– 30078 (cit. on p. 25)

    2020

  6. [6]

    YOLOv4: Optimal Speed and Accuracy of Object Detection

    A. Bochkovskiy, C.-Y. Wang, and H.-Y. M. Liao. “YOLOv4: Optimal Speed and Accuracy of Object Detection”. In: ArXiv abs/2004.10934 (2020) (cit. on pp. 47, 48, 151)

    arXiv 2020

  7. [7]

    Utilizing Amari-Alpha Divergence to Stabilize the Training of Generative Adversarial Networks

    L. Cai, Y. Chen, N. Cai, W. Cheng, and H. Wang. “Utilizing Amari-Alpha Divergence to Stabilize the Training of Generative Adversarial Networks”. In: Entropy 22 (2020) (cit. on pp. 42, 151)

    2020

  8. [8]

    Emerging Properties in Self-Supervised Vision Transformers

    M. Caron, H. Touvron, I. Misra, H. J’egou, J. Mairal, P. Bojanowski, and A. Joulin. “Emerging Properties in Self-Supervised Vision Transformers”. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV) (2021), pp. 9630–9640 (cit. on pp. 39, 41, 151). 138 BIBLIOGRAPHY

    2021

Show all 107 references
  1. [9]

    Chakraborty, A

    A. Chakraborty, A. Sharma, C. Pancholi, J. Arora, J. Grover, and P. Agar- wal. ANOMALY DETECTION IN VIDEO FEEDS. url: https:// ash- shar.github.io/reports/ML-Report-Anomaly-Detection.pdf (cit. on p. 15)

  2. [10]

    Anomaly detection: A survey

    V. Chandola, A. Banerjee, and V. Kumar. “Anomaly detection: A survey”. In: ACM Comput. Surv. 41 (2009), 15:1–15:58 (cit. on p. 15)

    2009

  3. [11]

    Grad-CAM++: Generalized Gradient-Based Visual Explanations for Deep Convolutional Networks

    A. Chattopadhyay, A. Sarkar, P. Howlader, and V. N. Balasubramanian. “Grad-CAM++: Generalized Gradient-Based Visual Explanations for Deep Convolutional Networks”. In: 2018 IEEE Winter Conference on Applica- tions of Computer Vision (W ACV) (2017), pp. 839–847 (cit. on p. 54)

    2017

  4. [12]

    You Only Look One-level Feature

    Q. Chen, Y. Wang, T. Yang, X. Zhang, J. Cheng, and J. Sun. “You Only Look One-level Feature”. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021), pp. 13034–13043 (cit. on pp. 49, 151)

    2021

  5. [13]

    Autoencoder-based network anomaly detection

    Z. Chen, C. Yeo, B.-S. Lee, and C. T. Lau. “Autoencoder-based network anomaly detection”. In: 2018 Wireless Telecommunications Symposium (WTS) (2018), pp. 1–5 (cit. on p. 34)

    2018

  6. [14]

    Xception: Deep Learning with Depthwise Separable Convolu- tions

    F. Chollet. “Xception: Deep Learning with Depthwise Separable Convolu- tions”. In: 2017 IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR) (2016), pp. 1800–1807 (cit. on p. 31)

    2016

  7. [15]

    Chollet et al

    F. Chollet et al. Keras. 2015. url: https://github.com/fchollet/keras (cit. on pp. 35, 55, 151)

    2015

  8. [16]

    Abnormal Event Detection in Videos us- ing Spatiotemporal Autoencoder

    Y. S. Chong and Y. H. Tay. “Abnormal Event Detection in Videos us- ing Spatiotemporal Autoencoder”. In: International Symposium on Neural Networks. 2017. doi: 10.1007/978-3-319-59081-3_23 (cit. on p. 34)

    2017 doi

  9. [17]

    Residual spatiotemporal autoencoder for unsupervised video anomaly detection

    K. Deepak, S. Chandrakala, and C. K. Mohan. “Residual spatiotemporal autoencoder for unsupervised video anomaly detection”. In: Signal, Image and Video Processing 15 (2020), pp. 215–222 (cit. on p. 34)

    2020

  10. [18]

    Adversarial autoencoders for anomalous event detection in images

    A. Dimokranitou. “Adversarial autoencoders for anomalous event detection in images”. PhD thesis. Purdue University, 2017 (cit. on p. 41)

    2017

  11. [19]

    Continual Learning for Anomaly Detection in Surveillance Videos

    K. Doshi and Y. Yilmaz. “Continual Learning for Anomaly Detection in Surveillance Videos”. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPR W) (2020), pp. 1025–1034 (cit. on pp. 17, 56). Fabien Poirier Paris 8 University BIBLIOGRAPHY 139

    2020

  12. [20]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Un- terthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale”. In: ArXiv abs/2010.11929 (2020) (cit....

    2020 arXiv

  13. [21]

    Finding Structure in Time

    J. L. Elman. “Finding Structure in Time”. In: Cogn. Sci. 14 (1990), pp. 179– 211 (cit. on p. 30)

    1990

  14. [22]

    You only look at one sequence: Rethinking transformer in vision through object detection

    Y. Fang, B. Liao, X. Wang, J. Fang, J. Qi, R. Wu, J. Niu, and W. Liu. “You only look at one sequence: Rethinking transformer in vision through object detection”. In: Advances in Neural Information Processing Systems 34 (2021), pp. 26183–26197 (cit. on pp. 50, 51, 151)

    2021

  15. [23]

    P. Ferlet. How to work with Time Distributed data in a neural network

  16. [24]

    P. Ferlet. keras-video-generators. 2020. url: https : / / github . com / metal3d/keras-video-generators

    2020

  17. [25]

    R. Gandhi. R-CNN, Fast R-CNN, Faster R-CNN, YOLO — Object Detec- tion Algorithms. 2018. url: https://towardsdatascience.com/r- cnn- fast - r - cnn - faster - r - cnn - yolo - object - detection - algorithms - 36d53571365e (cit. on p. 56)

    2018

  18. [26]

    YOLOX: Exceeding YOLO Series in 2021

    Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun. “YOLOX: Exceeding YOLO Series in 2021”. In: ArXiv abs/2107.08430 (2021) (cit. on p. 51)

    2021 arXiv

  19. [27]

    Learning to Forget: Contin- ual Prediction with LSTM

    F. A. Gers, J. Schmidhuber, and F. Cummins. “Learning to Forget: Contin- ual Prediction with LSTM”. In: Neural Computation 12 (2000), pp. 2451– 2471 (cit. on p. 30)

    2000

  20. [28]

    R. Ghosh. https://blog.qure.ai/notes/deep-learning-for-videos-action-recognition- review. 2018. url: https://blog.qure.ai/notes/deep-learning-for- videos-action-recognition-review (cit. on pp. 33, 42, 151)

    2018

  21. [29]

    Fast R-CNN

    R. B. Girshick. “Fast R-CNN”. In: 2015 IEEE International Conference on Computer Vision (ICCV) (2015), pp. 1440–1448 (cit. on p. 44)

    2015

  22. [30]

    Rich Feature Hierar- chies for Accurate Object Detection and Semantic Segmentation

    R. B. Girshick, J. Donahue, T. Darrell, and J. Malik. “Rich Feature Hierar- chies for Accurate Object Detection and Semantic Segmentation”. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition (2013), pp. 580–587 (cit. on pp. 43, 44, 151)

    2013

  23. [31]

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative Adversarial Networks. 2014. arXiv: 1406.2661 [stat.ML] (cit. on p. 41). Anomaly Detection in Videos 2023 140 BIBLIOGRAPHY

    2014 arXiv

  24. [32]

    M3d- CAM: A PyTorch library to generate 3D data attention maps for medical deep learning

    K. Gotkowski, C. Gonz´ alez, A. M. Bucher, and A. Mukhopadhyay. “M3d- CAM: A PyTorch library to generate 3D data attention maps for medical deep learning”. In: ArXiv abs/2007.00453 (2020) (cit. on p. 55)

    2020 arXiv

  25. [33]

    Learning Temporal Regularity in Video Sequences

    M. Hasan, J. Choi, J. Neumann, A. K. Roy-Chowdhury, and L. S. Davis. “Learning Temporal Regularity in Video Sequences”. In: 2016 IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR) (2016), pp. 733– 742 (cit. on p. 34)

    2016

  26. [34]

    Es- caping the Big Data Paradigm with Compact Transformers

    A. Hassani, S. Walton, N. Shah, A. Abuduweili, J. Li, and H. Shi. “Es- caping the Big Data Paradigm with Compact Transformers”. In: ArXiv abs/2104.05704 (2021) (cit. on pp. 40, 151)

    2021 arXiv

  27. [35]

    Deep Residual Learning for Image Recognition

    K. He, X. Zhang, S. Ren, and J. Sun. “Deep Residual Learning for Image Recognition”. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2015), pp. 770–778 (cit. on p. 31)

    2015

  28. [36]

    Long Short-Term Memory

    S. Hochreiter and J. Schmidhuber. “Long Short-Term Memory”. In: Neural Computation 9 (1997), pp. 1735–1780 (cit. on p. 30)

    1997

  29. [37]

    Densely Connected Convolu- tional Networks

    G. Huang, Z. Liu, and K. Q. Weinberger. “Densely Connected Convolu- tional Networks”. In: 2017 IEEE Conference on Computer Vision and Pat- tern Recognition (CVPR) (2016), pp. 2261–2269 (cit. on pp. 47, 151)

    2016

  30. [38]

    S. Jacob. Anomaly detection from videos: A deep learning approach. McGill University (Canada), 2019 (cit. on pp. 24, 30, 34)

    2019

  31. [39]

    Mobile Neural Architecture Search Net- work and Convolutional Long Short-Term Memory-Based Deep Features Toward Detecting Violence from Video

    H. M. B. Jahlan and L. A. Elrefaei. “Mobile Neural Architecture Search Net- work and Convolutional Long Short-Term Memory-Based Deep Features Toward Detecting Violence from Video”. In: Arabian Journal for Science and Engineering 46 (2021), pp. 8549–8563 (cit. on p. 32)

    2021

  32. [40]

    Incremental Training for Image Classification of Unseen Objects

    H. Jain and S. Nandy. “Incremental Training for Image Classification of Unseen Objects”. In: ResearchGate, Aug (2019) (cit. on pp. 46, 151)

    2019

  33. [41]

    3D Convolutional Neural Networks for Human Action Recognition

    S. Ji, W. Xu, M. Yang, and K. Yu. “3D Convolutional Neural Networks for Human Action Recognition”. In: IEEE Transactions on Pattern Analysis and Machine Intelligence 35 (2010), pp. 221–231 (cit. on pp. 32, 33, 151)

    2010

  34. [42]

    LayerCAM: Exploring Hierarchical Class Activation Maps for Localization

    P.-T. Jiang, C.-B. Zhang, Q. Hou, M.-M. Cheng, and Y. Wei. “LayerCAM: Exploring Hierarchical Class Activation Maps for Localization”. In: IEEE Transactions on Image Processing 30 (2021), pp. 5875–5888 (cit. on p. 54)

    2021

  35. [43]

    Serial Order: A Parallel Distributed Processing Approach

    M. I. Jordan. “Serial Order: A Parallel Distributed Processing Approach”. In: Advances in psychology 121 (1997), pp. 471–495 (cit. on p. 30)

    1997

  36. [44]

    Segment Anything

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y. Lo, P. Doll´ ar, and R. B. Girshick. “Segment Anything”. In: ArXiv abs/2304.02643 (2023) (cit. on p. 130). Fabien Poirier Paris 8 University BIBLIOGRAPHY 141

    2023 arXiv

  37. [45]

    Kotikalapudi and al

    R. Kotikalapudi and al. keras-vis. https://github.com/raghakot/keras- vis. 2017 (cit. on p. 54)

    2017

  38. [46]

    ImageNet classification with deep convolutional neural networks

    A. Krizhevsky, I. Sutskever, and G. E. Hinton. “ImageNet classification with deep convolutional neural networks”. In: Communications of the ACM 60 (2012), pp. 84–90 (cit. on p. 43)

    2012

  39. [47]

    Transfer Learning for Illustration Classifica- tion

    M. Lagunas and E. Garces. “Transfer Learning for Illustration Classifica- tion”. In: ArXiv abs/1806.02682 (2018) (cit. on p. 31)

    2018 arXiv

  40. [48]

    Temporal Convolutional Networks for Action Segmentation and Detection

    C. S. Lea, M. D. Flynn, R. Vidal, A. Reiter, and G. Hager. “Temporal Convolutional Networks for Action Segmentation and Detection”. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016), pp. 1003–1012 (cit. on pp. 35, 36, 151)

    2016

  41. [49]

    Backpropagation Applied to Handwritten Zip Code Recognition

    Y. LeCun, B. E. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. E. Hubbard, and L. D. Jackel. “Backpropagation Applied to Handwritten Zip Code Recognition”. In: Neural Computation 1 (1989), pp. 541–551 (cit. on p. 31)

    1989

  42. [50]

    Learning to detect anomaly events in crowd scenes from synthetic data

    W. Lin, J. Gao, Q. Wang, and X. Li. “Learning to detect anomaly events in crowd scenes from synthetic data”. In: Neurocomputing 436 (2021), pp. 248–259 (cit. on p. 41)

    2021

  43. [51]

    A Unified Approach to Interpreting Model Predictions

    S. M. Lundberg and S.-I. Lee. “A Unified Approach to Interpreting Model Predictions”. In: ArXiv abs/1705.07874 (2017) (cit. on p. 54)

    2017 arXiv

  44. [52]

    Remembering history with convolutional LSTM for anomaly detection

    W. Luo, W. Liu, and S. Gao. “Remembering history with convolutional LSTM for anomaly detection”. In: 2017 IEEE International Conference on Multimedia and Expo (ICME) (2017), pp. 439–444 (cit. on p. 34)

    2017

  45. [53]

    A motion-aware ConvLSTM network for action recognition

    M. Majd and R. Safabakhsh. “A motion-aware ConvLSTM network for action recognition”. In: Applied Intelligence 49 (2019), pp. 2515–2521 (cit. on p. 32)

    2019

  46. [54]

    Anomaly Detection in Video Using Pre- dictive Convolutional Long Short-Term Memory Networks

    J. R. Medel and A. E. Savakis. “Anomaly Detection in Video Using Pre- dictive Convolutional Long Short-Term Memory Networks”. In: ArXiv abs/1612.00390 (2016) (cit. on p. 34)

    2016 arXiv

  47. [55]

    A hybrid approach for search and rescue using 3DCNN and PSO

    B. Mishra, D. Garg, P. Narang, and V. K. Mishra. “A hybrid approach for search and rescue using 3DCNN and PSO”. In: Neural Computing and Applications (2020), pp. 1–15 (cit. on p. 32)

    2020

  48. [56]

    C. Molnar. Interpretable Machine Learning. A Guide for Making Black Box Models Explainable. 2019 (cit. on p. 55)

    2019

  49. [57]

    Real-Time Video Anomaly Detection for Smart Surveillance

    M. Mustafa. “Real-Time Video Anomaly Detection for Smart Surveillance”. In: SSRN Electronic Journal (2022) (cit. on p. 17). Anomaly Detection in Videos 2023 142 BIBLIOGRAPHY

    2022

  50. [58]

    Feature Visualization

    C. Olah, A. Mordvintsev, and L. Schubert. “Feature Visualization”. In: Dis- till (2017). https://distill.pub/2017/feature-visualization. doi: 10.23915/ distill.00007 (cit. on p. 25)

    2017

  51. [59]

    The Building Blocks of Interpretability

    C. Olah, A. Satyanarayan, I. Johnson, S. Carter, L. Schubert, K. Ye, and A. Mordvintsev. “The Building Blocks of Interpretability”. In: Distill (2018). https://distill.pub/2018/building-blocks. doi: 10 . 23915 / distill . 00010 (cit. on p. 25)

    2018

  52. [60]

    Temporal Fusion Approach for Video Classification with Convolutional and LSTM Neural Networks Applied to Violence Detection

    J. P. de Oliveira Lima and C. M. S. Figueiredo. “Temporal Fusion Approach for Video Classification with Convolutional and LSTM Neural Networks Applied to Violence Detection”. In: Inteligencia Artif. 24 (2021), pp. 40–50 (cit. on p. 32)

    2021

  53. [61]

    D´ etection d’anomalies en temps r´ eel dans le flux vid´ eo

    F. Poirier, R. Jaziri, C. Srour, and G. Bernard. “D´ etection d’anomalies en temps r´ eel dans le flux vid´ eo”. In: Extraction et Gestion des Connaissances: EGC’2022 38 (2022) (cit. on p. 8)

    2022

  54. [62]

    Enhancing Anomaly De- tection in Videos using a Combined YOLO and a VGG GRU Approach

    F. Poirier, R. Jaziri, C. Srour, and G. Bernard. “Enhancing Anomaly De- tection in Videos using a Combined YOLO and a VGG GRU Approach”. In: 2023 20th ACS/IEEE International Conference on Computer Systems and Applications (AICCSA). IEEE. 2023, pp. 1–6 (cit. on p. 8)

    2023

  55. [63]

    From CNN to ConvRNN: Adapting Visualization Techniques for Time-Series Anomaly Detection

    F. Poirier, R. Jaziri, C. Srour, and G. Bernard. “From CNN to ConvRNN: Adapting Visualization Techniques for Time-Series Anomaly Detection”. June 2023. url: https://arxiv.org/abs/2411.04707 (cit. on p. 8)

    2023 arXiv

  56. [64]

    Exploring Convolutional Recurrent architectures for anomaly detection in videos: a comparative study

    A. Ravi and F. Karray. “Exploring Convolutional Recurrent architectures for anomaly detection in videos: a comparative study”. In: Discover Artifi- cial Intelligence 1 (2021) (cit. on p. 32)

    2021

  57. [65]

    You Only Look Once: Unified, Real-Time Object Detection

    J. Redmon, S. K. Divvala, R. B. Girshick, and A. Farhadi. “You Only Look Once: Unified, Real-Time Object Detection”. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2015), pp. 779–788 (cit. on pp. 45, 151)

    2015

  58. [66]

    YOLO9000: Better, Faster, Stronger

    J. Redmon and A. Farhadi. “YOLO9000: Better, Faster, Stronger”. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016), pp. 6517–6525 (cit. on p. 46)

    2016

  59. [67]

    YOLOv3: An Incremental Improvement

    J. Redmon and A. Farhadi. “YOLOv3: An Incremental Improvement”. In: ArXiv abs/1804.02767 (2018) (cit. on p. 46)

    2018 arXiv

  60. [68]

    P. Remy. Keract: A library for visualizing activations and gradients. https: //github.com/philipperemy/keract. 2020 (cit. on p. 55). Fabien Poirier Paris 8 University BIBLIOGRAPHY 143

    2020

  61. [69]

    Faster R-CNN: Towards Real- Time Object Detection with Region Proposal Networks

    S. Ren, K. He, R. B. Girshick, and J. Sun. “Faster R-CNN: Towards Real- Time Object Detection with Region Proposal Networks”. In: IEEE Trans- actions on Pattern Analysis and Machine Intelligence 39 (2015), pp. 1137– 1149 (cit. on p. 44)

    2015

  62. [70]

    A study of deep convolu- tional auto-encoders for anomaly detection in videos

    M. Ribeiro, A. E. Lazzaretti, and H. S. Lopes. “A study of deep convolu- tional auto-encoders for anomaly detection in videos”. In: Pattern Recognit. Lett. 105 (2018), pp. 13–22 (cit. on p. 34)

    2018

  63. [71]

    “Why Should I Trust You?

    M. T. Ribeiro, S. Singh, and C. Guestrin. ““Why Should I Trust You?”: Explaining the Predictions of Any Classifier”. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2016) (cit. on p. 54)

    2016

  64. [72]

    Learning represen- tations by back-propagating errors

    D. E. Rumelhart, G. E. Hinton, and R. J. Williams. “Learning represen- tations by back-propagating errors”. In: Nature 323 (1986), pp. 533–536 (cit. on p. 34)

    1986

  65. [73]

    Unsupervised Anomaly Detection with Generative Adversarial Net- works to Guide Marker Discovery

    T. Schlegl, P. Seeb¨ ock, S. M. Waldstein, U. M. Schmidt-Erfurth, and G. Langs. “Unsupervised Anomaly Detection with Generative Adversarial Net- works to Guide Marker Discovery”. In: ArXiv abs/1703.05921 (2017) (cit. on p. 41)

    2017 arXiv

  66. [74]

    Grad-CAM: Visual Explanations from Deep Networks via Gradient- Based Localization

    R. R. Selvaraju, A. Das, R. Vedantam, M. Cogswell, D. Parikh, and D. Ba- tra. “Grad-CAM: Visual Explanations from Deep Networks via Gradient- Based Localization”. In: International Journal of Computer Vision 128 (2016), pp. 336–359 (cit. on p. 54)

    2016

  67. [75]

    Deep Learning for Automatic Violence Detection: Tests on the AIRTLab Dataset

    P. Sernani, N. Falcionelli, S. Tomassini, P. Contardo, and A. F. Dragoni. “Deep Learning for Automatic Violence Detection: Tests on the AIRTLab Dataset”. In: IEEE Access 9 (2021), pp. 160580–160595 (cit. on p. 23)

    2021

  68. [76]

    P. Sharma. A Practical Guide to Object Detection using the Popular YOLO Framework. 2018. url: https://www.analyticsvidhya.com/blog/2018/ 12/practical-guide-object-detection-yolo-framewor-python/ (cit. on p. 54)

    2018

  69. [77]

    P. Sharma. A Practical Implementation of the Faster R-CNN Algorithm for Object Detection. 2018. url: https://www.analyticsvidhya.com/blog/ 2018/11/implementation-faster-r-cnn-python-object-detection/ (cit. on p. 53)

    2018

  70. [78]

    P. Sharma. A Step-by-Step Introduction to the Basic Object Detection Algorithms. 2018. url: https://www.analyticsvidhya.com/blog/2018/ 10/a-step-by-step-introduction-to-the-basic-object-detection- algorithms-part-1/ (cit. on p. 53). Anomaly Detection in Videos 2023 144 BIBLIOGRAPHY

    2018

  71. [79]

    Convolutional LSTM Network: A Machine Learning Approach for Precip- itation Nowcasting

    X. Shi, Z. Chen, H. Wang, D. Y. Yeung, W.-K. Wong, and W.-c. Woo. “Convolutional LSTM Network: A Machine Learning Approach for Precip- itation Nowcasting”. In: NIPS. 2015 (cit. on p. 32)

    2015

  72. [80]

    Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps

    K. Simonyan, A. Vedaldi, and A. Zisserman. “Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps”. In: CoRR abs/1312.6034 (2013) (cit. on p. 54)

    2013 arXiv

  73. [81]

    Very Deep Convolutional Networks for Large-Scale Image Recognition

    K. Simonyan and A. Zisserman. “Very Deep Convolutional Networks for Large-Scale Image Recognition”. In: CoRR abs/1409.1556 (2014) (cit. on p. 31)

    2014 arXiv

  74. [82]

    Smooth- Grad: removing noise by adding noise

    D. Smilkov, N. Thorat, B. Kim, F. B. Vi´ egas, and M. Wattenberg. “Smooth- Grad: removing noise by adding noise”. In: ArXiv abs/1706.03825 (2017) (cit. on p. 54)

    2017 arXiv

  75. [83]

    L’apprentissage non-supervis´ e et ses contradictions

    J. Sublime. “L’apprentissage non-supervis´ e et ses contradictions”. In: Bul- letin 1024 (2022) (cit. on pp. 34, 151)

    2022

  76. [84]

    Real-world anomaly detection in surveillance videos

    W. Sultani, C. Chen, and M. Shah. “Real-world anomaly detection in surveillance videos”. In: Proceedings of the IEEE conference on computer vision and pattern recognition. 2018, pp. 6479–6488 (cit. on p. 20)

    2018

  77. [85]

    Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning

    C. Szegedy, S. Ioffe, V. Vanhoucke, and A. A. Alemi. “Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning”. In: ArXiv abs/1602.07261 (2016) (cit. on p. 31)

    2016 arXiv

  78. [86]

    Going deeper with convolutions

    C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. E. Reed, D. Anguelov, D. Er- han, V. Vanhoucke, and A. Rabinovich. “Going deeper with convolutions”. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2014), pp. 1–9 (cit. on p. 31)

    2014

  79. [87]

    EfficientNet: Rethinking Model Scaling for Convo- lutional Neural Networks

    M. Tan and Q. V. Le. “EfficientNet: Rethinking Model Scaling for Convo- lutional Neural Networks”. In: ArXiv abs/1905.11946 (2019) (cit. on p. 31)

    2019 arXiv

  80. [88]

    Employing long short-term memory and Facebook prophet model in air temperature forecasting

    T. Toharudin, R. S. Pontoh, R. E. Caraka, S. Zahroh, Y. Lee, and R. C. Chen. “Employing long short-term memory and Facebook prophet model in air temperature forecasting”. In: Communications in Statistics - Simulation and Computation 52 (2021), pp. 279–290 (cit. on pp. 30, 151)

    2021

  81. [89]

    Learning Spatiotemporal Features with 3D Convolutional Networks

    D. Tran, L. D. Bourdev, R. Fergus, L. Torresani, and M. Paluri. “Learning Spatiotemporal Features with 3D Convolutional Networks”. In: 2015 IEEE International Conference on Computer Vision (ICCV) (2014), pp. 4489– 4497 (cit. on p. 32)

    2014

  82. [90]

    Selective Search for Object Recognition

    J. R. R. Uijlings, K. E. A. van de Sande, T. Gevers, and A. W. M. Smeul- ders. “Selective Search for Object Recognition”. In: International Journal of Computer Vision 104 (2013), pp. 154–171 (cit. on p. 44). Fabien Poirier Paris 8 University BIBLIOGRAPHY 145

    2013

  83. [91]

    Attention is all you need

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. “Attention is all you need”. In: Advances in neural information processing systems 30 (2017) (cit. on pp. 36, 38, 151)

    2017

  84. [92]

    Rapid object detection using a boosted cas- cade of simple features

    P. A. Viola and M. J. Jones. “Rapid object detection using a boosted cas- cade of simple features”. In: Proceedings of the 2001 IEEE Computer Soci- ety Conference on Computer Vision and Pattern Recognition. CVPR 2001 1 (2001), pp. I–I (cit. on pp. 42, 43, 151)

    2001

  85. [93]

    A New Approach for Abnormal Human Activities Recognition Based on ConvLSTM Architec- ture

    R. Vrskova, R. Hudec, P. Kamencay, and P. Sykora. “A New Approach for Abnormal Human Activities Recognition Based on ConvLSTM Architec- ture”. In: Sensors (Basel, Switzerland) 22 (2022) (cit. on pp. 23, 32)

    2022

  86. [94]

    Human Activity Clas- sification Using the 3DCNN Architecture

    R. Vrskova, R. Hudec, P. Kamencay, and P. Sykora. “Human Activity Clas- sification Using the 3DCNN Architecture”. In: Applied Sciences (2022) (cit. on p. 32)

    2022

  87. [95]

    Violent Be- havioral Activity Classification Using Artificial Neural Network

    R. Vrskova, R. Hudec, P. Sykora, P. Kamencay, and M. Benco. “Violent Be- havioral Activity Classification Using Artificial Neural Network”. In: 2020 New Trends in Signal Processing (NTSP) (2020), pp. 1–5 (cit. on p. 32)

    2020

  88. [96]

    YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors

    C.-Y. Wang, A. Bochkovskiy, and H.-Y. M. Liao. “YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors”. In: ArXiv abs/2207.02696 (2022) (cit. on pp. 52, 53, 152)

    2022 arXiv

  89. [97]

    CSPNet: A New Backbone that can Enhance Learning Capability of CNN

    C.-Y. Wang, H.-Y. M. Liao, I.-H. Yeh, Y.-H. Wu, P.-Y. Chen, and J.-W. Hsieh. “CSPNet: A New Backbone that can Enhance Learning Capability of CNN”. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPR W) (2019), pp. 1571–1580 (cit. on pp. 48, 151)

    2019

  90. [98]

    You Only Learn One Representation: Unified Network for Multiple Tasks

    C.-Y. Wang, I.-H. Yeh, and H. Liao. “You Only Learn One Representation: Unified Network for Multiple Tasks”. In: J. Inf. Sci. Eng. 39 (2021), pp. 691– 709 (cit. on pp. 50, 151)

    2021

  91. [99]

    Score-CAM: Score-Weighted Visual Explanations for Convolutional Neural Networks

    H. Wang, Z. Wang, M. Du, F. Yang, Z. Zhang, S. Ding, P. ( Mardziel, and X. Hu. “Score-CAM: Score-Weighted Visual Explanations for Convolutional Neural Networks”. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPR W) (2019), pp. 111–119 (cit...

    2019

  92. [100]

    Abnormal Event Detection in Videos Using Hybrid Spatio-Temporal Autoencoder

    L. Wang, F. Zhou, Z. Li, W. Zuo, and H. Tan. “Abnormal Event Detection in Videos Using Hybrid Spatio-Temporal Autoencoder”. In: 2018 25th IEEE International Conference on Image Processing (ICIP) (2018), pp. 2276–2280 (cit. on p. 34)

    2018

  93. [101]

    P. Wang. Vision Transformer - Pytorch Overview. 2020. url: https : / / github.com/lucidrains/vit-pytorch (cit. on pp. 40, 41, 151). Anomaly Detection in Videos 2023 146 BIBLIOGRAPHY

    2020

  94. [102]

    Visual Transformers: Token-based Image Representation and Processing for Computer Vision

    B. Wu, C. Xu, X. Dai, A. Wan, P. Zhang, M. Tomizuka, K. Keutzer, and P. Vajda. “Visual Transformers: Token-based Image Representation and Processing for Computer Vision”. In: ArXiv abs/2006.03677 (2020) (cit. on pp. 41, 151)

    2020 arXiv

  95. [103]

    YOLOP: You Only Look Once for Panoptic Driving Perception

    D. Wu, M. Liao, W.-T. Zhang, X. Wang, X. Bai, W.-Q. Cheng, and W.-Y. Liu. “YOLOP: You Only Look Once for Panoptic Driving Perception”. In: Machine Intelligence Research 19 (2021), pp. 550–562 (cit. on pp. 49, 151)

    2021

  96. [104]

    Interpretable Convolutional Neu- ral Networks

    Q. Zhang, Y. N. Wu, and S.-C. Zhu. “Interpretable Convolutional Neu- ral Networks”. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (2017), pp. 8827–8836 (cit. on p. 25)

    2017

  97. [105]

    Video Anomaly Detection for Smart Surveillance

    S. Zhu, C. Chen, and W. Sultani. “Video Anomaly Detection for Smart Surveillance”. In: ArXiv abs/2004.00222 (2020) (cit. on pp. 16, 19, 42, 147)

    2020 arXiv

  98. [106]

    Object Detection in 20 Years: A Survey

    Z. Zou, Z. Shi, Y. Guo, and J. Ye. “Object Detection in 20 Years: A Survey”. In: Proceedings of the IEEE 111 (2019), pp. 257–276 (cit. on p. 54). Fabien Poirier Paris 8 University List of Tables 1.1 Datasets mentioned by S. Zhu, C. Chen, and Sultani (2020) . . . . 19 1.2 Categ...

    2019

  99. [2019]

    url: https://medium.com/smileinnovation/how-to-work-with- time-distributed-data-in-a-neural-network-b8b39aa4ce00 (cit. on pp. 77, 152)

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.