Abnormal behavior detection using sparse representations through sequential generalization of k-means
The potential capability to automatically detect and classify human behavior as either normal or abnormal events is an important aspect in intelligent monitoring/surveillance systems. This study presents a new high-performance framework for detecting behavioral abnormalities in video streams by util...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Turkiye Klinikleri
2021
|
| Subjects: | |
| Online Access: | http://eprints.utm.my/id/eprint/94025/1/AhlamAlDhamari2021_AbnormalBehaviorDetectionUsingSparse.pdf http://eprints.utm.my/id/eprint/94025/ http://dx.doi.org/10.3906/ELK-1904-187 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Universiti Teknologi Malaysia |
| Language: | English |
| Summary: | The potential capability to automatically detect and classify human behavior as either normal or abnormal events is an important aspect in intelligent monitoring/surveillance systems. This study presents a new high-performance framework for detecting behavioral abnormalities in video streams by utilizing only the patterns for normal behaviors. In this paper, we used a hybrid descriptor, called a foreground optical flow energy (FGOFE), which makes use of two effective motion techniques in order to extract the most descriptive spatiotemporal features in video sequences. The FGOFE descriptor can effectively capture both weak and sudden incidents in a scene. The sequential generalization of k-means (SGK) algorithm was applied in this study to generate the dictionary set that can sparsely represent each signal; in addition, the orthogonal matching pursuit algorithm was utilized to recover high-dimensional sparse features when referring to a few numbers of noisy linear measurements. Using the SGK allows gaining a less complex and quicker implementation compared to other dictionary learning methods. We conducted comprehensive experiments to analyze and evaluate the ability of our framework in detecting abnormalities using several public benchmarks, which contain different abnormal samples and various contextual compositions. The experimental results show that the proposed framework achieved high detection accuracy (up to 95.33%) and low frame processing time (31 ms on average) compared to the relevant related work. |
|---|