Accurate Football Detection and Localization for Nao Robot with the Improved HOG-SVM Approach
Zejun Bi, Qi Yang, Mengyi Zhang, Alban Goupil, Lihang Feng · 2018
Under the background of the RoboCup, this essay aims to solve the problem of finding and localizing the football in the limited computation hardware environment. Machine learning method has been used in the present study, for its much better robustness comparing with the traditional methods, such as template matching and contour detection. In this paper, we use hough circle transform for detection preliminarily. After this operation, only a few candidates will be left. Then we extract candidates' histogram of oriented gradient, using support vector machine (SVM) to predict if the candidate contains football. Relying on the successful detection, we construct a space model to retrieve the distance. In experiment section, the precision and recall were calculated on a test set, which indicates our method is accurate. Also, we compared the ground truth of distance with theoretical value, finding the error is negligible.