XNOR-YOLO: The High Precision of The Ball and Goal Detecting on The Barelang-FC Robot Soccer

Abstract

The essential part in developing the humanoid robot which is able to play soccer is the vision system. The vision system needs to be fast and accurate in detecting the surrounding objects on the field such as the ball, goal, teammates, or even the opponent. One of the powerful methods which was able to generate the object detection quickly with high accuracy was the deep learning. However, this method proceeded a huge computation. Even if it was generated on the GPU, it would still generat a low speed of detecting. Therefore, the high precision and fast detecting of the object method need to be considered in this area. In order to overcome this problem, we proposed the combination of the XNOR-Network (XNOR-Net) towards YOLOv3 running on the GPU with the same layer configuration as the tinyYOLO. To testify the performance of this method, some experiments has been carried out in real-time application by implementing it in the NVDIA Jetson TX1 GPU. From the experiment results, this method is able to detect the object faster than other object detection in detecting the ball and goal colored by white and generated 30 FPS in detecting each object.