Research On Object Recognition And Localization For Middle Size League Soccer Robot Under Variant Illumination

The robot soccer is a new interdisciplinary research area,referring to diverse fields like intelligent robotics, multi-agent system, real-time image processing, pattern recognition, path planning, machine learning, etc. It provides a standard testing platform for evaluation of Artificial Intelligence theories and robotics. Middle-Size League, as a crucial competition in RoboCup, is the most technologically challenging. The autonomous mobile robot is provided with the abilities of self-organization, self-layout and adaptability. It is dependent on color information obtained from the color camera for their operation, such as object recognition, object localization, path planning, kicking, shooting, etc. However, the color of the object is sensitive to illumination changes. The pixel values represented by an image of an object will be different to another image of the same object taken under different lighting conditions. For a non-uniform lighting, some object recognition algorithms based on color will fail. The illumination has become the bottleneck of the object recognition and localization for soccer robot.In this dissertation, object recognition and localization for soccer robot under variant illumination are studied. The main achievements of research work are done:(1) The light source, illumination model, theories of color vision and model of color camera imaging is addressed. The color image segmentation, object recognition and localization are introduced. The characteristic of the object recognition and localization for the soccer robot is analyzed. The algorithms of object recognition and localization for the soccer robot under variant illumination are surveyed, and the advantages and disadvantages of these algorithms are analyzed. Furthermore, the trend of object recognition and localization for the soccer robot is described.(2) A method of on-line adaptation to illumination is proposed for mobile robot in changing illumination environment. Illumination condition is represented by an average luminance distribution of green pixels in a time series images. Illumination is classified into different levels from the bright to dark condition. Color calibration is done under each illumination condition. Illumination change is detected by computing the KL-divergence between two different distributions. A dual-threshold strategy is used to classify the current illumination into known conditions or an unknown one. According to illumination the robot decides to switch to a corresponding color calibration or learn a new one. Experimental results show the efficiency of the proposed method.(3) A robust method of objects recognition for autonomous soccer robot based on support vector machine and Gabor filters is proposed. Firstly, candidate targets are extracted based on color threshold and simple shape features. Then, images containing candidate targets are convoluted with the Gabor filter, from which the feature vector is extracted. The target recognition is performed by the support vector machine. Experiments have been carried out on the soccer robot M-TR for the ball recognition. Experimental results prove the high target recognition accuracy and real-time performance of the proposed method.(4) Target localization for soccer robot based on Omni-directional vision is considered. The vision system of the MT-R robot is described. A fast method of object localization based on piece-wise defined Lagrange interpolation is proposed. The polar coordinate transforming is adopted based on the imaging characteristic of the Omni-directional vision system. The localization formula is established based on distance calibration. A distance look-up table is set up to create a mapping from the pixel radius in the image and actual distance radius on the field between the object and the robot. Experimental results show the effectiveness of the proposed localization method.