Research On Key Techniques Of High Efficient Video Object Tracking Under Complex Environments

Video object tracking is one of the key tasks in computer vision system, which has a wide application foreground in many fields such as Intelligent Video Surveillance, Human-Computer Interaction, Medical Diagnosis and Robot Navigation. However, video object tracking still faces many challenges from many factors in the actual application environments, such as illumination change, shadow, occlusion, abrupt motion, background clutter, etc.. This dissertation deeply studies the key technologies of high efficient video object tracking, and the main research contents and relevant achievements in this dissertation are as follows.(1) Aiming at the problem of the background-pixel disturbance in a tracked object region more or less under complex environments, the object tracking method based on foreground hue histogram is proposed. This method constructs the object model combining the result of motion segmentation and the feature of hue histogram, and the vector of the foreground hue histogram about object model is effectively updated during the tracking process. This method effectively enhances the tracking stability when the background-pixels are included in a rectangular tracked region under complex scenes.(2) Aiming at the problem of the tracking failure from noise disturbance or error matching via a single SIFT feature under complex environments, the object tracking method combining SIFT feature and hue histogram feature is proposed. This method constructs the object model fusing SIFT feature and hue histogram feature based on a fair condition by using a normalized measure. The object model is updated during the tracking process. This method obtains a better tracking stability under complex scenes than that based on a single SIFT feature.(3) Aiming at the problem of effectively combining compressive sensing theory and video motion object tracking process, the tracking model based on particle filter via compressive sensing is constructed, and three worth problems are discussed, which are selecting compressive measurement matrix, realizing compressive measurement process, and analyzing compressive measurement vector. Under constraint of this model, the compressive tracking method based on particle filter is proposed. This method compressively samples Haar-like feature vector based on compressive sensing theory and predicts candidate objects based on particle filter algorithm. Comparing with the existed compressive tracking methods, this method enhances the tracking real-time and retains the tracking stability, which realizes the high efficiency video object tracking.(4) Aiming at the problem of selecting compressive measurement matrix, the faster compressive tracking method based on adaptive measurement matrix is proposed. The sparse degree, the column number and the row number of this measurement matrix adaptively changes with a given tracked object, and the measurement matrix can preserve the distances between the points in an original signal vector space with a less error and with high probability. Comparing with the newly existed tracking methods based on Haar-like feature, a more stable and real-time effect can be obtained under complex scenes, and this method further realizes the higher efficiency video object tracking. Meanwhile, there gives three experimental phenomena about selecting the measurement matrix. Firstly, the lower sparse measurement matrix can cause that every measurement element includes more original information. Secondly, the computing complexity of compressively measuring process rests with two aspects:the sparse degree of measurement matrix and the number of measurement elements. Thirdlly, the success rate of tracking process has not a corresponding relation with the number of measurement elements.