Research On The Visual Target Tracking

With the rapid development of modern computer science and informationtechnology, and with the marvelous improvement of the algorithms of imageidentification, visual target tracking already becomes a challenging research topic inmachine vision at present, which integrates advanced technologies in many fields suchas computer vision, image processing, pattern recognition, artificial intelligence andautomatic control. It has a lot of potential applications in national defence, securitysurveillance, intelligent traffic system, etc.The main focuses of the thesis are to improve the visual target tracking systemframework—particle filtering algorithm. The particle filter integrates the colorinformation and the motion information when computing the importance weights, so asto enhance the accuracy of visual target tracking. The thesis firstly introduces severalcommon algorithms in visual object detection, and makes an improvement; Secondly, itintroduces some common-used filtering techniques such as the Kalman filter, theGauss-Hermite filter, the particle filter and the Extended filtering techniques on thebasis. The Unscented Kalman filter and the particle filter will be emphasized; Thirdly, aparticle filtering algorithm based on the Gauss-Hermite filter and the Unscented Kalmanfilter is proposed. In this new algorithm, it integrates the Gauss-Hermite filter and theUnscented Kalman filter to generate the importance density function. By do this,problems such as particle degeneracy problem which is caused by the general particlefilter using transition prior density function as proposal distribution are solved. At last, anew visual target tracking algorithm, whose framework is particle filter, is proposed.This algorithm integrates the color information and the motion information whencomputing the importance weights. The simulation results show that: the new particlefiltering algorithm outperforms obviously the unscented particle filtering algorithm, andthe RMSE decreases by 70%; At the same time, the new visual target tracking algorithmcan improve the robustness and the accuracy.