Laser & Optoelectronics Progress, Volume. 58, Issue 6, 600004(2021)
Research on Video Abnormal Behavior Detection Based on Deep Learning
Figures & Tables(5)
Table 1. Advantages and disadvantages of the 3 abnormal behavior detection algorithms
View tableTable 1. Advantages and disadvantages of the 3 abnormal behavior detection algorithms
Algorithm Advantage Disadvantage Supervised learning relatively less training samples is needed; the most accurate; easy to understand and apply accurate label is time-consuming; only predefined anomaly can be detected; hard to generalize Weakly supervised learning relatively accurate with weak label; lower false alarm; relatively easy to use split the difference; relatively lower accuracy Unsupervised learning training without any label; only normal data needed; robust and easy to generalize higher false positive rate; poor positioning accuracy; unable to classify anomaly behavior
Table 2. Comparison of commonly used anomaly behavior detection datasets
View tableTable 2. Comparison of commonly used anomaly behavior detection datasets
Dataset Video Length Annotation Anomaly category UCSD 98 10 min pixel-level biker, skater, wheelchair, car, walking on the grass Avenue 37 30 min frame-level run, throw, abnormal object UMN 5 5 min frame-level group escape UCF-Crime 1900 128 h video-level abuse, arrest, arson, assault, accident, burglary, fighting, robbery
Table 3. Performance comparison of different algorithms unit: %
View tableTable 3. Performance comparison of different algorithms unit: %
Algorithm UCSD Ped1 UCSD Ped2 Avenue Frame-level Pixel-level Frame-level Pixel-level Frame-level EER AUC EER AUC EER AUC EER AUC EER AUC DSTCNN[ 12 ]-- 99.74 -- -- -- 99.94 -- -- -- -- LDA-Net[ 13 ]-- -- -- -- 5.63 97.87 12.91 92.96 -- -- FCN+LSTM[ 17 ]-- -- -- -- -- -- 6.6 98.2 -- -- TW-C3D[ 18 ]6.29 96.73 9.22 95.27 5.59 96.37 11.80 93.51 -- -- MISVM[ 31 ]22 -- -- -- 16 -- -- -- 21 84.5 Algorithm UCSD Ped1 UCSD Ped2 Avenue Frame-level Pixel-level Frame-level Pixel-level Frame-level EER AUC EER AUC EER AUC EER AUC EER AUC MLEP[ 32 ]-- -- -- -- -- -- -- -- 24.8 92.8 AMDN[ 34 ]16.0 92.1 40.1 67.2 17.0 90.8 -- -- -- -- OC SVM[ 35 ]-- -- -- -- 10.6 93 17.3 88 -- -- GMFC-VAE[ 36 ]11.3 94.9 36.3 71.4 12.6 92.2 19.2 78.2 22.7 83.4 MGFC-AAE[ 37 ]20 85 -- 72.6 16 91.6 -- 88 22.3 84.2 CAE[ 39 ]-- 89.5 -- -- -- 54.7 -- -- -- 75.4 STAE[ 40 ]15.3 92.3 -- -- 16.7 91.2 -- -- 24.4 80.9 LSTM AE[ 43 ]12.5 89.9 -- -- 12.0 87.4 -- -- -- 80.3 3D-LSTM AE[ 44 ]15.9 90.9 -- -- 15.8 93.6 -- -- 20.7 81.8 S2-VAE[ 45 ]14.3 -- -- 94.25 11.54 95.77 14.28 90.83 -- 87.6 GAN[ 48 ]8 97.4 35 70.3 14 93.5 -- -- -- -- Predict-GAN[ 51 ]-- 83.1 -- -- -- 95.4 -- -- -- 84.9 ST U-Net[ 52 ]22.3 83.82 -- -- 8.7 96.56 -- -- -- 84.59 Bi-Prediction[ 53 ]-- 89.0 -- -- -- 96.6 -- -- 21.5 87.8
Tools
Get Citation
Copy Citation Text
Peng Jiali, Zhao Yingliang, Wang Liming. Research on Video Abnormal Behavior Detection Based on Deep Learning[J]. Laser & Optoelectronics Progress, 2021, 58(6): 600004
Paper Information
Category: Reviews
Received: Jun. 19, 2020
Accepted: --
Published Online: Mar. 6, 2021
The Author Email: Yingliang Zhao (zhaoyl18@nuc.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20