Improvement And Implementation Of Object Tracking Algorithm In Intelligent Video Analysis

Today, Intelligent Video Analysis(IVA) plays an important role in the era of big data world. As one of Key techniques in IVA, object tracking is of great academic value and wide range of applications. The paper focuses on the improvement and implementation of object tracking algorithm, firstly discusses and analysises the classical object tracking algorithm--Mean Shift and its drawbacks,then introduces the improved ideas and finally implements and optimizes the algorithm on hardware system.Firstly, the intelligent video analysis system based on TMS320DM8168 is developed, which support up to 4-way video processing and SD/HD output. Meanwhile, the intelligent video object tracking software framework based on McFW is built, which can process single or multiple target.Secondly, the classic object tracking algorithm--Mean Shift(MS) was learned. As MS is small amount of calculation and easy to implement, it became the best choice from similation platform to hardware implementation.The classic MS tracking algorithm only uses color as the target description, so it is difficult to handle the similar color between a target and another target or background, and even worse, it may lost the object. To overcome this problem, the paper gives multi-feature fusion tracking algorithm combined the texture and color in two ways, then achieves the tracking algorithm in MATLAB. The multi-feature fusion tracking algorithm is evaluated on a variety of scenarios videos compared with traditional mean shift algorithm based on color. Simulation results demonstrate that the multi-feature fusion tracking algorithm had better performance for the color similarity problems. Besides, it performs more robust on illumination variation, occlusion and fast motion.Thirdly, the factors impacting the robust object tracking in the long time were analysised and discussed, then improved strategies combined with motion detection were suggested. The traditional mean shift tracking algorithm always selects a rectangle to represent the target model, which inevitably contains background pixels that may impact the accuracy of the target location. So the paper suggested a background suppression target model way, which introducing motion detection to avoid background pixels, to maximise differentiation of target and background for accurately tracking. Meanwhile, in the long-term tracking, the object may change its scale and pose, occur occlusion or disappear from views. So the paper combined the motion detection with object tracking for updating target template timely and detecting the reappear in the views. These strategies make sense to improve the accuracy and robustness in long-term MS tracking.Lastly, the improved mean shift tracking algorithm based on multi-feature fusion of texture and color is implemented and optimized on the DM8168. While ensuring the accuracy, tracking algorithm is optimized as much as possible. At last, the tracking speed is up to 13 ms per frame, which achieve real-time processing. To make the video object tracking system more intelligent, the paper gives single target tracking and multitarget tracking framework on DM8168 hardware system, then the system can track multiple target in the scene. In a word, the research of this paper could provide some help for subsequent behavior recognition and other applications.