Classification of Moving Crowd Based on Motion Pattern

Crowd behavior analysis is a significant task in the context of surveillance and crowd management. For a moving crowd, analyzing the motion pattern is very important. In this work, we present a simple scheme to categorize such crowds as structured, semi-structured and unstructured ones. The categorization is achieved based on the regularity of the motion pattern of the collection of objects (humans, in this case). In case of structured one, the movement is coherent and uniform in nature. It is expected that the crowd as a whole or individual segment of it reflects consistent orientation and speed of movement. For unstructured crowd, on the other hand, the movement is random. Hence, diversity is there in terms of orientation and speed. The semi-structured one stands in between and makes the classification problem difficult. In this work motion orientation based feature is computed to represent the motion pattern. A set of interest points detected in the initial frame are tracked over the sequence using optical flow. Thus, motion orientations are obtained. A frame is divided into blocks, and distribution of the orientation of motion of the interest points in each block is summarized in a four dimensional histogram. Block level histograms are concatenated to form the frame level descriptor. Finally, frame level descriptors are taken together to represent the sequence. In this experiment, artificial neural network (ANN) is used as classifier. Experiment is carried out on collectiveness dataset. Proposed method provides better classification accuracy in comparison to state-of-the-art techniques.