The Research Of Object Detection And Tracking Based On Fixed Single Viewpoint

Moving object detection and tracking technology is widely used in intelligent video surveillance, human-computer interaction, military application fields etc. It's also the foundation of target identification, classification and behavior analysis and understanding. It's one of the hotspots in the field of computer vision. This thesis tries to get insights on some key issues, such as background updating for foreground detection, objects with obvious scale-changes, occlusion for objects tracking and so on. The main research contents are as follows:1) This thesis proposed a new algorithm using Time Information Window-Kernel Density Estimation (TIW-KDE) for background dirt, large amount of calculation in the background updating phase of the classical non-parametric kernel density estimation. This algorithm, which took full advantage of the information on the foreground frames along the time line, divided the background into dynamic background region and non-dynamic background region. For the dynamic background region, the algorithm used non-parametric kernel density estimation algorithm to update it, otherwise, the percent of background and current frame was used to progressively update the non-dynamic background region. This effectively settled the problems of imprecise foreground object detection and lower real-time.2) For the particle filtering algorithm based on color feature can not effectively deal with the targets, which with obvious size-changes in tracking process, this thesis proposed a new automatic tracking window scale updating algorithm. This algorithm extracted the primal sketch of object based on the visual theory, and then used the changes of the elements-number of the primal sketch as the measure information to judge the situation of target's size-changes, which was then used to improve the particle filtering algorithm based on color histogram. This effectively settled the problem of adjust the tracking window scale adaptively in the tracking process.3) This thesis proposed an improved particle filtering algorithm, which combined with sub-blocks matching and trajectory prediction for the traditional particle filtering algorithm can not effectively deal with object under occlusions in the tracking process. This algorithm judged the situation of the target's occlusion firstly, when the target was occluded partly, it divided the target template into some sub-blocks, and then estimated the target location through the best matching blocks searched by each sub-block; When the target was occluded seriously, it predicted the target's location by it's historical trajectory. This new algorithm effectively enhanced the particle filtering algorithm's robustness in dealing with the problem of occlusion.