Research On Adaptive Object Recognition Applied To Omni-directional Vision System Of Soccer Robot

As a symbolic subject of artificial intelligence and robotics, robot caught more and more researchers attention. It has been widely used in industry, agriculture, military, medicine and other fields. Medium-sized robot soccer is a complex multi-agent system. It not only refers to communication technology, machine vision, computer graphics, artificial intelligence, but also involves the technology of robotics, mechatronics, sensotics, intelligent control etc.. Among them, the ability to recognize objecrt of robot soccer vision system is the building blocks of robot path planning, motion control, and cooperation to complete the task, The omni-directional vision system is the most prominent in vision system, so how to improve the object recognition accuracy and light adaptive ability of the robot omni-directional vision system is the main issue discussed in this thesis.According to problem that robot soccer environment easily affected by the changing illumination and lead to the low object recognition rate and the poor light adaptability, the main research are as follows:(1)The image enhancement algorithm of soccer robot omni-directional vision system.Briefly expounds the traditional homomorphic filter algorithm and analyzed it’s advantages and disadvantages. On the basis of the traditional homomorphic filter algorithm, using wavelet transform to replace fourier transform for image enhancement. Due to the illumination change mainly reflects in luminance component I of the HSI color space, we only need to deal with I. After logarithm transform with I, we use Mallat wavelet decomposed it into low frequency and high frequency. Then use butterworth homomorphic filtering function to filtering the wavelet coefficient, and refactor the filtered wavelet and inverse logarithm transformate it and last have the enhanced image.(2)A new object recognition algorithm based on biSCAN and SVM. Designed a new color model IV which is specifically for target orange ball. We use the within class scatter matrix and the between class scatter matrix to chose two components I and V from the six components of HSI and normalization YUV. I and V is the most sensitive to the orange color of the ball. Uses biSCAN to set a few radial scanning which start from the image center and along the radius to extract object area, and then training support vector machine (SVM) to recognize the object based on IV color model, so as to improve the accuracy of object recognition and adaptive ability to light.At the end of this thesis, it summarizes the main research of this task, also points out the shortcomings and the problem which needs further study in the future.