Video Based Motion Trajectory Analysis of Pedestrians

Abstract

Video cameras are widely used for surveillance in public domains, which allow security personnel to remotely monitor the activities in an area of interest. Due to inadequate manpower, effective manual monitoring of multiple video feeds is not possible. With progress in video processing and computing technologies, it is possible to develop real time algorithms to intelligently detect and decipher scenes in the video frames. In public domains like bus and rail interchanges, shopping malls, numerous applications can be derived from identifying pedestrian crowds. Pedestrian crowds can be analyzed by tracking the movement of individuals to understand the crowd behavior. Automatic group identification and tracking can highlight regions of interest from numerous video feeds, thereby aiding the security personnel in the surveillance of the area.

In this thesis, a three-level blob filter utilizing the pedestrian detections from Histogram of Oriented Gradients for human detection and Background Subtraction is developed to detect pedestrians. The three-level blob filter addresses pedestrian detection challenges such as illumination variations and pedestrian-like confusers.

A method to identify social groups of pedestrians from pedestrian crowd videos is proposed. The method identifies social groups based on the social psychological theory of pedestrian groups, by monitoring whether the pedestrians move closely together in the same direction with almost the same velocity for a certain period of time. The pedestrian group features such as group size, group member locations, time of group identification are extracted and stored in a proposed data structure, termed as Pedestrian Group Record. Several applications which use the information in the Pedestrian Group Record are proposed. They are Real-time pedestrian meeting-event detection system, Stall pedestrian occupancy determination, People Counting at queuing regions and door locations, Crowd density estimation at train platforms and Visual representations of pedestrian’s group history. The proposed methods and the applications are validated using real-world video data sets. Some of the applications are deployed in real-world locations after the validation.

Pedestrian motion changes which happen during the convergence of pedestrians to form a group is utilized to predict future pedestrian groups. In this proposed method, the pair wise pedestrian motion changes which lead to group formations are learnt by a Support Vector Machine. The learnt model is utilized to predict new pairwise pedestrian motion change. A similar approach is proposed to predict approach of potential customers to food outlets.