The Algorithm Researches On Multi-Objects Detection And Tracking In Video

Object tracking has become an important research field in several domains such as image processing, pattern recognition and computer vision. It prevails in diverse applications in visual surveillance, intelligent traffic and military affairs. This dissertation accomplishes automatic detection and pedestrians tracking based on the analysis of video sequences. To deal with the pedestrians tracking problem in video surveillance, an algorithm based on Gaussian probability is proposed. Modeling object on the basis of statistical characteristics of color information which is grouped by improved k-means algorithm, it segments object into blocks according to the grouping results, and utilizes a Gaussian function in order to describe the distribution of color information in each block as well as the position information of pedestrians. The algorithm accomplishes when the maximum joint probability of color and position is obtained. Along with position information and color information of objects among adjacent frames, the failure case which is because of information loss due to occlusion could be solved. The multi-cameras are used to extend the range of surveillance. For multi-object tracking problem, the same pedestrian in different camera should be verified using the geometric information between cameras. The paper proposes an approach which adopts 3D information, and the improvement of the approach which uses epipolar line constraint and histogram matching. The multi-cameras should be calibrated before the geometric information is imported. In order to overcome the disadvantage in current algorithms, a new algorithm called RANSAC external parameters calculation based on videos is developed. While people are walking around, feature point correspondences could be extracted under control. By using the accumulation of correspondences along the sequence, this algorithm removed the outliers and end up with an accurate result. Not only RANSAC but also the dynamic effect of feature points is used to get an accurate result. Finally, the algorithm is tested with videos taken in outdoor environment, and the results are provided in this dissertation.