Research On Video Based Object Detection And Tracking

The technology of video based object detection and tracking is one of the hotspots in the field of computer vision, which is also the basic and important technology in the applications of smart surveillance, human-machine interface, mobile robots navigation, industrial robots hand-eye system and so on. This thesis focuses on the solution to the problems of the objects, especially people, the detection and the tracking in the applications of smart surveillance and human-machine interface. Main achievements of this dissertation are as follows:1. An approach to detect and track moving objects in a static background sequence was proposed. Firstly, a GMM model of the gray distribution of a difference image was constructed and a motion detection operator was introduced to generate a motion border image. Secondly, GVF-Snake was enhanced to extract the contours of the moving objects by modifying the energy entry and adding automatic snake initialization. The contour position at next moment was predicted to accelerate convergence of Snake. Based on the above mentioned algorithm, normalized RGB space combined with gray space model was proposed to replace the plain gray space model so as to immune the effect of shadows.2. Faced with the problem that most present pedestrian detecting algorithms can not deal with multi-view objects and can not perform in real time, a fast multi-view pedestrian detecting algorithm was provided. Two detectors were utilized with one being based on skin and hair color and the other on head and shoulder contours. In addition, a coarse to fine matching algorithm was introduced to detect head-shoulder contours faster. Finally, a cascaded detecting system was constructed by the two detectors. The system can run at a maximum speed of 30fps with 352×288 resolution.3. A particle filter based tracking algorithm was also developed to track the contour of a pedestrian, which can deal with nonlinearity caused by clutters and decrease the computing complexity due to the using of the same feature as above.4. Mean Shift based tracking algorithm can perform well in term of translationtracking but can not deal with the rotation tracking, so a Mean Shift based rotation tracking algorithm was proposed, which utilized the gray gradient direction distribution of the target region as the feature and constructed a similarity function that can be optimized by Mean Shift method, thus the rotation tracking was transformed into the optimization problem. Due to the fast convergence of Mean Shift, this algorithm can run in real-time. A complete algorithm that can track both translation and rotation of targets was obtained by alternate iteration of rotation and translation tracking.