Research And Applications Of Video Object Tracking

Intelligent video surveillance is one of the most important research domains in computer vision, and plays a key role in security protection and military protection. Intelligent video surveillance aims to achieve automatic detection, tracking and recognition of objects using image processing, machine learning and computer vision technologies. Moreover, it will help to analyze and understand the behavior of moving objects. In intelligent video surveillance, object tracking lends itself as the basis of object recognition and behavior analysis, and its performance influences the whole system directly. Therefore, research of object tracking is of great significance in theory and application.To date, many researchers from both academia and industry have made great efforts and proposed a lot of valuable object tracking algorithms. Of them, Mean Shift and Particle Filter are two of the most mature and useful in practice. However, there still exist some drawbacks for these two algorithms. For example, they may not work well in cases that background is similar to target, or there is severe partial occlusion, which are prevalent in practical applications. Consequently, it is still difficult to design practical and effective object tracking algorithm. Based on the proposed algorithm, this dissertation studies the object tracking algorithm, and makes contributions as follows:Mean Shift object tracking algorithm is investigated and its disadvantages are analyzed. Then, a rapid and efficient object tracking algorithm-MAP Spatial Pyramid(MAP-SP)Mean Shift is proposed in this dissertation. The proposed algorithm considers the background information into Mean Shift framework and divides the target dynamically in the tracking process, to adaptively keep geometric structure.Local feature based object tracking algorithm is also investigated. A new local feature descriptor is proposed to avoid the high complexity of traditional local feature descriptor SIFT. This new descriptor is introduced into the MAP-SP Mean Shift framework to improve tracking performance. The proposed descriptor is simple and easy to implement. Therefore, it will improve the performance of tracking in real-time demand.Experimental results demonstrate that the proposed approaches can overcome some drawbacks of existing algorithms, satisfy real-time demand and improve the performance of tracking.This dissertation designs and completes an experimental video surveillance system based on the proposed algorithms. The system adopts the module design, consisting of motion detection, object analysis and object tracking. The efficiency of this system is demonstrated via comparative experiments on both standard and our own video sequences, providing an experimental platform for the latter research.