Research Of Moving Target Detection And Tracking

Moving target detection and tracking is one of the most important subjects in computer vision, which combines advanced technologies in image processing, pattern recognition, artificial intelligence, automatic control, computer and other relative fields. It has very important practical value and wide developmental prosperity in military missile guidance, visual navigation, video surveillance, intelligent transportation, medical image analysis, industrial product detection, etc.This thesis mainly aims to study algorithms of detecting and tracking moving target in stationary background. Firstly, the moving target detection method based on Gaussian Mixture Model is discussed. Then, two detection algorithms are proposed. One is based on background reconstruction, and the other is based on removing background edges and GVF (Gradient Vector Flow) Snake. Finally, the target tracking method based on mean shift is studied, and an improved tracking method which combines mean shift and adaptive prediction is proposed to solve the existing problems. The major contributions of this thesis are summarized as follows:In the moving target detection part, the detection method based on Gaussian Mixture Model is studied. Considering it has poor adaptability to abrupt illumination change, an improved detection method is proposed, which integrates the whole illumination change in judging whether the current pixel value matches the background model. Considering current detection methods based on background reconstruction usually have problems of requiring heavy computation or large memory, a new target detection method based on background reconstruction is proposed. Firstly, the background is reconstructed based on the improved symmetrical difference method and block operation. Then, the moving target is detected based on background subtraction. The difference threshold is automatically selected by two-dimensional Otsu's method. The new method requires less computation, need not store history images, and has some adaptability to the scene change. Considering background subtraction methods based on region operation usually have the problems of requiring exact background models and being sensitive to illumination change, a detection method based on removing background edges and GVF Snake is proposed, where the target edges are extracted by the edge-based background removal method, the target close contour is gained by GVF Snake, and the whole target region is detected by morphological filling process. These problems that background subtraction methods based on region operation have are solved in a certain extent by this method.In the moving target tracking part, the tracking method based on mean shift is studied, which uses the kernel-weighted color histogram to characterize the target. This method has good adaptability to target round, transmogrification, and part-occlusion. Furthermore, it requires simple computation and has real-time performance. However, it can not adjust the size of the tracking window adaptively, and has poor performance in tracking fast moving targets and targets with maneuverable movement. A tracking method based on mean shift and adaptive prediction is proposed to improve the performance of the current mean shift tracker. Firstly, the target model is constructed combining target shape information and color discriminability. Then, the state of the target is estimated according to the matching degree. When there is no maneuverable movement and occlusion, the mean shift tracker with kalman prediction is used to achieve real-time performance in tracking fast moving target. Otherwise, the mean shift tracker with particle prediction is used to achieve robust tracking performance. Few particles with two different motion-models are used to further improve the robustness and real-time performance of the tracker. The speediness of the mean shift tracker with kalman filter and the robustness of the mean shift tracker with particle filter are integrated efficiently in the improved tracking method, which achieves real-time and robust tracking performance.