Precise Extraction of Partially Occluded Objects by Using HLAC Features and SVM

In the RoboCup competition, robot soccer game, ball and robots are extracted by using color information. If color markers attached on the robot or a ball itself are occluded, especially the occlusion ratio is high, it will be difficult to extract them. This paper proposes a new and high precision method which extracts partially occluded objects based on the statistical features of the pixel and its neighborhoods. Concretely, at first, input image is labeled by using color information and small candidate regions which have similar color to the color markers or the ball are extracted, then each candidate region is classified into partially occluded object or noise by using HLAC features and SVM. We applied our method to the global vision system of RoboCup small size league (SSL) and confirmed that it could extract partially occluded objects, 94.23% for 5 to 8 pixels area and 80.06% for 3 to 4 pixels area, and worked more than 60fps.