Trajectory Clustering in Surveillance Videos Using Dynamic Time Warping

Abstract

Adding functional analysis to the videos captured by surveillance cameras can provide handy information to their users. Mining the existing trajectories in a video is one of the most valuable features, discovering the prevalent patterns and their density in the video, and it helps reveal some unusual and abnormal movements more easily. In this paper, the data obtained through the execution of detection and tracking algorithms are processed in various steps and used to train a hierarchical clustering model by deploying a modified version of the DTW algorithm. This practical approach does not need massive datasets for the training procedure and can be applied to any surveillance video containing different types of objects. The proposed method utilizes information extracted from the objects in a video to generate the existing primary trajectories. Additionally, a practical algorithm for modeling the background in surveillance movies is proposed to illustrate clustering outputs.