Multi-Pedestrians Anomaly Detection via Conditional Random Field and Deep Learning

Abstract

Automated video surveillance frameworks quickly distinguish surprising and basic circumstances in a packed climate that would help to pursue sufficient choices for security and crisis control. Hence, In this paper, an innovative method for automatically detect and localize anomalous objects among multi-pedestrian crowds via conditional random field and deep learning is introduced. Initially, necessary preprocessing is performed on extracted frames and then super-pixels are generated using improved watershed transform, the objects are then segmented using a conditional random field. The region of interests are localized using conditional probability and temporal association is implemented to locate the regions with a group of pedestrians and pedestrians with other objects. A deep learning feature pyramid network is then implemented to detect and categorized the objects in each region and finally, the anomalous objects are identified using Jaccard similarity. The effectiveness of proposed framework is assessed on openly accessible UCSD Ped 1 and Ped 2 datasets and it accomplishes an accuracy rate of 94.2% and 95.4% respectively. Extensive experimental data and comparative analysis show that our model outperformed current state-of-the-art models in terms of accuracy.