Motion anomaly detection and trajectory analysis in visual surveillance
Manaswi Chebiyyam,Rohit Desam Reddy,Debi Prosad Dogra,Harish Bhaskar,Lyudmila Mihaylova
DOI: https://doi.org/10.1007/s11042-017-5196-6
IF: 2.577
2017-09-30
Multimedia Tools and Applications
Abstract:Motion anomaly detection through video analysis is important for delivering autonomous situation awareness in public places. Surveillance scene segmentation and representation is the preliminary step to implementation anomaly detection. Surveillance scene can be represented using Region Association Graph (RAG), where nodes represent regions and edges denote connectivity among the regions. Existing RAG-based analysis algorithms assume simple anomalies such as moving objects visit statistically unimportant or abandoned regions. However, complex anomalies such as an object encircles within a particular region (Type-I) or within a set of regions (Type-II). In this paper, we extract statistical features from a given set of object trajectories and train multi-class support vector machines (SVM) to deal with each type of anomaly. In the testing phase, a given test trajectory is categorized as normal or anomalous with respect to the trained models. Performance evaluation of the proposed algorithm has been carried out on public as well as our own datasets. We have recorded sensitivity as high as 86% and fall-out rate as low as 9% in experimental evaluation of the proposed technique. We have carried out comparative analysis with state-of-the-art techniques to benchmark the method. It has been observed that the proposed model is consistent and highly accurate across challenging datasets.
computer science, information systems, theory & methods,engineering, electrical & electronic, software engineering