Study On Moving Object Detection And Tracking Under Complex Scenes

Moving object detection and tracking is a hot topic in the computer vision and pattern recognition, and playing an important role in intelligent surveillance, human-computer interaction, visual navigation and so on. Although numerous visual object detection and tracking algorithms have been proposed, there still exist a number of difficulties in real applications, such as complex background and object appearance changes, which seriously affect the robustness of detection and tracking algorithms. In this paper, we focus on moving object detection and tracking methods under the complex environment. The contribution of the thesis can be summarized as follows:(1) Aiming at the difficulties in background subtraction algorithm which is sensitive to moving object exist in the first frame, dynamic background and illumination variation, a novel background subtraction algorithm based on spatio-temporal sample consensus(STSC) is proposed. First, a sample-based spatio-temporal background model is presented, we model each background pixel by a history of recently observed pixel values and spatial neighborhood pixel values in 3×3×n video patches. Second, in order to suppress the interference of dynamic background, we propose the concept of background sample spatio-temporal entropy, which can segment the image into static region(i.e. single modal region) and dynamic region(i.e. muti-modal region), then use the dynamic threshold method to classify the truth foreground pixels. Third, in order to make the background model can better adapt to the changes of background and the moving object, we combines the background pixel-level update method with foreground blob-level update method, and put forward the two-level update strategy. Finally, in order to eliminate the influence of illumination variation, a brightness transform method is proposed. Sufficient experiments and comparisons with several state-of-the-art algorithms show the effectiveness and superiority of our algorithm.(2) For the shortcoming that the color feature is sensitive to illumination, an improved particle filter based on color feature for object tracking algorithm is proposed. First we improved the traditional histogram weighted function by using a scale factor. Then, a new color local entropy object observation model is constructed by mapping the object from color feature space to local entropy space. In addition, in order to make the algorithm adjust to the object deformation and environmental interference better, an adaptive updating strategy of the object template is designed and the number of particle can be adjusted dynamically according to the tracking performance. The experimental results show that compared with several existing algorithms, the proposed algorithm is more effective and robust for the real-time object tracking under the condition of illumination variation, object occlusion and non-linear motion.(3) In order to improve the robustness for long time object tracking, we introduce the object detection into the online learning object tracking algorithm, an online learning algorithm for object tracking based on improved compressive tracking and object detection is proposed, which contains three components: detector, tracker and tracking state analyzer. The proposed algorithm detect and track the object independently, tracking state analyzer is used to evaluate the result of detector and tracker, then determine the current tracking status and formulate the corresponding processing method, and updating the parameters adaptively according to the tracking result. Sufficient experiments and comparisons with several state-of-the-art algorithms show the effectiveness and superiority of our algorithm.(4) Aiming at the difficulty in the multi-object tracking system: data association and object interaction problem, a multi-object tracking algorithm based on data association and particle filter is proposed. First STSC background subtraction algorithm is used to detect the moving objects. Second the foreground objects are associated with the existing objects by using the two-level association matching method. Third determined the status of each object and give the corresponding solutions by analyzing correlation matrix, when object interaction happens, use particle filter to tracking objects. Finally, online learning and updating the objects feature adaptively according to the result of data association. Sufficient experiments show the effectiveness and superiority of our algorithm.