Crowd Escape Event Detection via Pooling Features of Optical Flow for Intelligent Video Surveillance Systems

In this paper we propose a method for automatic detection of crowd escape behaviour. Motion features are extracted by optical flow using LucasKanade derivative of Gaussian method (LKDoG) followed by robust probabilistic weighted feature pooling operation. Probabilistic feature polling chooses the most descriptive features in the sub-block and summarizes the joint representation of the selected features by Probabilistic Weighted Optical Flow Magnitude Histogram (PWOFMH) and Probabilistic Weighted Optical Flow Direction Histogram (PWOFDH). One class Extreme Learning Machine (OC-ELM) is used to train and test our proposed algorithm. The accuracy of our proposed method is evaluated on UMN, PETS 2009 and AVANUE datasets and correlations with the best in class techniques approves the upsides of our proposed method.