The Contour Tracking And Recognition Of Moving Objects In Vedios

The detection and tracking of moving objects is the core problem in computer vision field. It is a new technology which combines image processing, pattern recognition, artificial intelligence and other technologies together and it also has important practicality values and bright future in video surveillance, military vision control, medical diagnosis, controlled transportation and so on. The contour tracking can not only get the position of moving objects, but also obtain their contours which are the most important characters in the recognition of objects and their actions. Therefore, the contour tracking can be more helpful for the perception of objects and their actions.As part of the“video surveillance system”project, this dissertation mainly concentrates on the contour tracking and recognition of moving objects in surveillance videos. The main contents are as follows:1 This dissertation introduces the frameworks of video surveillance system, concludes the currently used object detection and tracking methods systematically by references to the latest literature and explicates Mean-Shift and particle filter which are the most popular tracking methods.2 This dissertation proposes a contour tracking algorithm based on the Condensation and Mean-Shift methods. It imports the Mean-Shift method into the framework of Condensation to remove its two defects: the inability to adjust of the parameters in states-shifting equation and the singles of the rules used to evaluate the particles. This algorithm combines Condensation method and Mean-Shift method organically and can track the objects better. 3 This dissertation proposes a contour tracking algorithm based on the Mean-Shift method and edge detection. This algorithm gets the center of the object using Mean-Shift tracking method firstly, then gets the foreground rectangle and background rectangle from the current frame and the updated background frame respectively, and then gets the contour of the object by background subtraction and edge detection method. It can track the object stably, accurately and timely.4 This dissertation proposes an object recognition algorithm based on the wavelet descriptor. It firstly describes the object contour using wavelet descriptor, then matches the wavelet descriptor of the unknown object with the ones of the sample objects and finds the most matched sample object which belongs to the same class with the unknown object. The wavelet descriptor is invariant to shape translation, scaling and rotation, so the algorithm has good performance.5 This dissertation also establishes a shape-based image classification system. This system can evaluate the performances of all kinds of shape descriptors. We can both add new descriptors or sample images into it and delete unused descriptors or sample images from it easily. This system can be used to compare the performances of all descriptors and can be a helpful experiment platform for the future research in object recognition.