EXPLOITING LOCAL INFORMATION FOR TRAJECTORY CLASSIFICATION UNDER SURVEILLANCE

Abstract

Object motion trajectory classification is the important task, especially when we aim to detect abnormal movement patterns in order to take appropriate actions for prohibiting unwanted events to occur. Given a set of trajectories recorded over a period of time, they can be clustered for understanding usual flow of movement or segmentation of unusual flows. With the advancement in low-cost sensors object trajectories can be recorded efficiently with ease. These sensors can be RGB video camera, depth camera like Kinect, and Global Positioning System (GPS). With the use of GPS, the real world coordinates of objects along with other related information are tracked and later processed for the real-time analysis of mass flow, crowd analysis, and anomaly detection in the flow. Video based trajectory analysis could be on-line or off-line. In on-line, objects are tracked in the live videos and there motion is analyzed immediately to make the higher order decisions like prohibiting the objects to enter restricted area, unstable areas like fire and floods, and situations like violence. Video trajectory classification is also done off-line, where object trajectories are first extracted from the recorded videos. Next, their motion is analyzed for off-line analysis by classifying the trajectories into different classes. In this thesis, we have focused on off-line analysis for the classification of object trajectories using the publicly available datasets. Using the local information along with global information is an efficient way to improve classification performance. To compute the local cues from trajectories, models could be built by partitioning the trajectories into variable number of segments based on the geometry of the trajectories. A graph based method for trajectory classification has been proposed. Each trajectory has been partitioned into varying number of segments based on its geometry. Complete Bipartite Graph (CBG) is formed between each trajectory pair and there Dynamic Time Warping (DTW) distance is used as the weight of the edge between them. Local costs are computed from the CBG and then fused (using Particle Swarm Optimization (PSO)) with the global cost (global cost computed using iii the same full length trajectory-pair) to improve the classification performance. We have also proposed a kernel transformation followed by trajectory classification framework that make the use of information from local segments. The proposed kernel perform the shrinking of trajectories in such a way they preserve their shape. Modified trajectories have been segmented with the help of segmental HMM and their local responses have been recorded. These local responses along with global responses (from full length trajectories) have been fused using genetic algorithm to make the final decision. A surveillance scene segmentation has been performed based on the results of trajectory classification using HMM. The scene layout is divided into 10 10 local non-overlapping grids and majority voting based scheme is applied to assign each block a label showing the importance of the blocks with the help of region association graph based features. Such off-line analysis helps to understand the flow of motion within the viewing field of video camera.