Improved abnormal detection using self-adaptive social force model for visual surveillance
With the growth of technology in computer vision, there is a great demand for an automated surveillance system in replaced to the traditional visual surveillance. The automated surveillance system is a system that monitors the behavior and activities of the crowd whether it is normal or not. The ab...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2017
|
| Subjects: | |
| Online Access: | http://umpir.ump.edu.my/id/eprint/19601/19/Improved%20abnormal%20detection%20using%20self-adaptive%20social%20force%20model%20for%20visual%20surveillance.pdf http://umpir.ump.edu.my/id/eprint/19601/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | With the growth of technology in computer vision, there is a great demand for an automated surveillance system in replaced to the traditional visual surveillance. The
automated surveillance system is a system that monitors the behavior and activities of the crowd whether it is normal or not. The abnormal detection in a crowd is a
noteworthy research topic in automated surveillance system in public places. It is emergent to detect the abnormal events as quickly as possible and take appropriate actions to minimize the loss and ensure the public safety. In this work, we aim to find the significant interaction forces and detect the abnormality in the crowd by using Self-Adaptive Social Force Model. For this point, Horn-Schunck optical flow is used to get
the flow vector for each pixel in the image frames. Instead of tracking individuals, particle advection is performed to capture continuity of crowd flow and its trajectories. These particles are then advected to a new location according to its underlying optical flow vector at the current location. Using the attained flow vectors from this stage, interaction force estimation is done based on SFM theory. This experiment is done with
the hypothesis that high magnitude of interaction force portrayed the abnormal behavior in a crowd. However, there is a problem with the earlier SFM, which is the similarity of actual velocity and desired velocity caused the abnormal detection inaccurate. The estimation of the good quality of interaction forces is critical in this case and has not been explored yet. So, Self-Adaptive SFM is developed in order to estimate a good quality of interaction forces since it is crucial to achieve better abnormal detection, which represents the behavior of the crowd. From the experiment, the highest and least magnitude of interaction force can be localized in the image frame. The proposed algorithm is validated with three challenging datasets contain abnormal videos,
including the videos of crime in Malaysia. For both indoor and outdoor scene, the proposed algorithm outperforms the other methods with accuracy 97% and 100%. For the benchmarking datasets, the AUC (Area under Curve) score of the proposed algorithm is quite comparable with previous works with the score of 0.9916. The AUC score provided by the proposed algorithm on PETS2009 datasets is about 0.9026 and 0.9940 for Malaysia Crime dataset. Based on these results, it can conclude that the high
magnitudes of interaction forces portray the abnormality in the scene and Self-Adaptive SFM is well-performed on crime scene with the rapid motion characteristic. |
|---|