Author Affiliations
1 Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang, Liaoning 1100162 University of Chinese Academy of Sciences, Beijing 1000493 Key Laboratory of Opto-Electronic Information Processing, Chinese Academy of Sciences, Shenyang, Liaoning 1100164 Key Laboratory of Image Understanding and Computer Vision, Shenyang, Liaoning 1100165 Space Star Technology Co., Ltd., Beijing 100086show less
Fig. 1. Flow chart of abnormal behavior detection and location
Fig. 2. Segmented motion-region images. (a) Original image; (b) motion regions achieved by the ViBE method; (c) image with motion magnitude; (d) motion regions achieved by the proposed algorithm
Fig. 3. Filtering results according to motion continuity for the abnormal regions which are predicted by the classifiers
Fig. 4. Examples of normal and abnormal crowed activities in UMN dataset. (a) Normal behaviors; (b) abnormal behaviors
Fig. 5. ROC curves of different methods in two criterions on the UCSD ped2 dataset. (a) Frame-level; (b) pixel-level
Fig. 6. Pixel-level detection results with different methods on the UCSD ped2 dataset. (a) Social Force; (b) MPPCA; (c) T-MDT; (d) S-MDT; (e) proposed method
Fig. 7. ROC curves of different methods for frame-level detection on the UMN dataset
Fig. 8. Detection and location results of three different scenes on UMN dataset
Fig. 9. Relationship curves between the parameter k and the system performance on two datasets
Method | Frame-level criterion | Pixel-level criterion | |
---|
EER /% | | AUC | DR /% | AUC | Time /s |
---|
Social Force[33] | 42.0 | 0.702 | 27.6 | 0.217 | - | MPPCA[34] | 31.1 | 0.710 | 22.4 | 0.222 | - | S-MDT[32] | 28.7 | 0.750 | 63.4 | 0.665 | 0.69 | T-MDT[32] | 27.9 | 0.765 | 56.8 | 0.522 | 0.64 | ST-CNN[23] | 24.4 | 0.860 | 81.9 | 0.880 | 0.37 | Motion Energy[16] | 22.0 | 0.810 | 55.0 | 0.580 | 0.08 | Proposed | 10.3 | 0.905 | 89.7 | 0.902 | 0.26 |
|
Table 1. Comparison results of different methods on the UCSD ped2 dataset for frame-level and pixel-level detection
Method | AUC | EER /% |
---|
Social Force[33] | 0.949 | 12.6 | Sparse[1] | 0.996 | 2.8 | H-MDT CRF[19] | 0.995 | 3.7 | ST-CNN[23] | 0.996 | 3.3 | Motion Energy[16] | 0.989 | 4.1 | Proposed | 0.989 | 2.7 |
|
Table 2. Comparison results of different methods on the UMN dataset in AUC and EER criterion for frame-level detection