Vision system is a very important sensor for robots to perceive surrounding environment. Since the visual information processing is usually time-consuming, it is necessary to develop a powerful vision system with sufficient computational resources and make the visual processing algorithms run on it, so that the workload of robot main processor is relived and the real-time performance of robots is improved. Because there are some special requirements on cost, size and power of robot vision systems for mobile robots applications, building a vision system based on embedded technology has become a research focus recently. One of the key points to apply embedded vision systems in practice is to improve the real-time performance of vision processing. Supported by National Natural Science Funds and 863 National High-Tech Research and Development Plan, we developed a DSP-based embedded vision system for mobile robots and studied several vision processing algorithms in this thesis. The main work and contributions of the thesis include following issues: 1. Design and implementation of a DSP-based embedded vision system for mobile robot are proposed. The developed system characterizes compact size, low power cost, high computational performance and excellent extensibility. Using an appropriative video/image digital signal processor, the system can perform vision processing really fast. Modular structure on both software and hardware and sufficient peripheral interfaces make it easy to be applied to robots with different requirements by little redesign. 2. A fast skin detection algorithm based on piecewise linearization method is proposed which makes use of the clustering properties of human skin in YCbCr color space. The proposed algorithm makes skin detection quite reliable since the influences which illuminance imposes on chrominance are taken into consideration. 3. A real-time face tracking method based on MeanShift iteration is proposed. This developed method extracts distributions of human face based on the proposed skin detection algorithm in advance, and then it makes use of the nonparametric MeanShift technique for climbing density gradients to find the mode of probability distributions of human face. 4. A particle filter tracking algorithm based on color distributions and similarity measurements is designed and implemented. The developed algorithm gets approximations of the object's position, size and motion parameters by Bayesian recursive estimation. It can perform robust racking of non-grid objects even in case of clutter and occlusion.
修改评论