A Fine Three-dimensional Adaptive Color Segmentation Algorithm for Real-time Vision System of Soccer Robot

The aim of the vision system for small soccer robots is to recognize and track robots and soccer ball in a semi-open environment. Traditional color segmentation methods are usually sensitive to external conditions, for example light changes or uneven distribution. Machine learning-based methods usually require advanced hardware as well as huge amount of data. To overcome these limitations, we propose a fine three-dimensional adaptive color classification method in this paper. The presented method determines the thresholds in color space dynamically. For distinguishing the overlapping area in color space, Naive Bayes classifier is used to classify the color categories which obey the Multivariate Normal Distribution. Dynamic adaptive lighting is introduced to enhance the performance, and the robustness of classification is increased by combining multiple brightness separable color space. This method is embedded in a vision system for Soccer Robot to enhance the processing performance. The experiments show that the system is robust and achieves high precision. It can process the soccer video in real-time with detection rate of 99.68 % and precision of 99.92 %.