Mean Shift Algorithm-based Video Vehicle Tracking

With the rapid development of intelligent transportation, the video surveillance system as its important component has been achieved much attention. In recent years, computer technology increasingly upgrading, and digital image process technology continues to mature, these technologies provide an effective technical support for the design of the video surveillance system. As the core technology of video surveillance, the video target's tracking and recognition become a very hot research topic. Video surveillance is a wide used, such as underwater detection, supermarkets'video surveillance, car navigation, traffic video surveillance. Video target tracking technology combines the advanced knowledge in many fields, including computer, image processing, communication system, algorithm research and so on, its prospects bright.Filter design is the core part of the video tracking, which is extract real data, models and other information from the noise environment, and these information can reflect the true state of the system. In the second part, this paper discussed the basic principles of linear filtering, then introduced the Kalman Filter design and it's development process. Kalman Filter is widely applied in communication or signal processing taking an irreplaceable role in many fields. However, in video target tracking the Mean Shift algorithm is much better than Kalman Filter, Mean Shift algorithm is based on the non-parameter density estimation, and it has a good adaptability in computer vision especially in the video tracking, so this paper focused on Mean Shift algorithm and elaborate its application in video tracking. In the third part of this paper, it described the Mean Shift derivation in detail on the multidimensional space and discussed its convergence. Meanwhile in the fourth part, it was compared with the particle filter and doing the comparison experiments between particle filter and Mean Shift, which completed in the same conditions and the results showed that the Mean Shift algorithm has great advantages.As the experiments of the third part shows that nuclear-width of Mean Shift algorithm remains unchanged for the variable size tracking targets will results tracking error. To overcome this defect, in the fifth part, the paper adapts Lindeberg's theory about feature scale selection to the problem of selecting kernel scale for mean-shift blob tracking. The new method can automatically choose an appropriate scale for blob tracking, and experiments showed that the improved algorithm can better adapt to tracking the blobs that are changing in size compared with the original algorithm.