Research On Visual Object Detecting And Tracking In Complex Environment

The objective of computer vision is the using of computer to take the place of the human eyes and brain with its ability to interprete the objects and the circumstance from the images that received. The techniques of object detecting and tracking are the focal points in the fileds of computer vision, object recognition, pattern analysis, video coding, content-based indexing and smart video surveillance. And these techniques have promising future either in scientific research or applications.The surveillance of dynamic circumstance is an interesting branch which attracts many attentions in recent years. Those are very important techniques of smart video surveillance to detect and track the moving objects in realistic complex circumstance. And the detecting and tracking result will influence the consequsent behaviors understanding. Some methods of visual objects detecting and tracking are proposed in this dissertation based on the theory of digital video and image processing, statistics, and dynamic system analysis. The research work may be summarized as follows.A novel objects detecting method is proposed based on the Gaussian mixture models and background subtraction and jointly weighting of hue and value components in HSV color space. The Gaussian mixture models are built up in HSV color space and the different states of each pixel are described in different Gaussian model. The order of the Gaussian models is sorted from the most likely one to the least. The parameters of Gaussian mixture models are updated according to the scenes changes. The hue and value component are jointly investigated to detect the moving foreground. The experimental results show that the proposed method is effective in complex circumstance such as wobbling branchs, waving surfaces, rain and snow.The object's shadows are eliminated in HSV color space by uniformly judgement of wether saturation and value are below certain thresholds and hue is similar. The saturation and value component of each pixel are compared to the same position in background model. If the saturation and value are below the specified threshould the pixel is probably in shadows. Then, if the hue component of suspectable shadow is the same as in backgroundm, the pixel will be classified into shadow rather than into moving object.A new tracking method for object in occlusion or not is proposed to improve the Mean Shift tracking algorithm with the Kalman filter. The Kalman filter predictor can reduce the Mean Shift iteration circums and handle the occlusion problem. The improved tracking method is more robust and can fit the request of real-time tracking. When the tracked object is in occlusion, the Kalman filter predicts the possible position of the object. And when the object is visible, the Mean Shift iteration starts from the Kalman filter prediction result. This algorithm can improve the accuracy and speed of moving object tracking. The experiment results show the improved tracking method's quickness and robustness.An improved particle filter method for multiple visual objects tracking is proposed with a Markov Chain Monte Carlo method and Mean Shift to solve problem of particles degeneration. The multiple objects tracking is a challenging problem due to the similar appearance of the the objects, background clutter, object merging and splitting, etc. The states of the objects are estimated using sequential Monte Carlo simulation method to overcome the limitation of the Kalman filter for nonlinear and/or non-Gaussian conditions. A Markov Chain Monte Carlo method and Mean Shift algorithm are introduced to solve problem of particles degeneration and clustering the particles.At the end of this dissertation, a smart video surveillance system is presented which employ the object detecting and tracking methods that proposed above. The feasibility to implement the proposed objects detecting and tracking algorithms in DSP in studied and a novel distributed smart video surveillance system architecture is proposed. The object detecting and tracking methods are tested in some realistic applications and results imply that those methods will be effective in the smart video surveillance system.