Object detection based on GPU parallel computing for RoboCup Middle Size League

The RoboCup Middle Size League (MSL) robot soccer competition is a standard test platform for distributed multi-robot systems. There are many challenges in the vision system for MSL soccer robots. For example, huge amount of data from the Kinect v2 sensor leads to heavy computation burden for the robot's onboard industrial computer, the obstacle-detection algorithm is mainly dependent on the obstacle' colors, the omnidirectional vision system is not able to detect the ball above the camera and get the objects' height information. In this paper, we proposed an algorithm for object detection based on GPU parallel computing employing Kinect v2 and Jetson TX1 as the hardware platform. Parallel computing is utilized throughout all the steps of the object detection algorithm, so the speed and accuracy of the algorithm are greatly improved. We test the real-time performance and the accuracy of the algorithm using our NuBot soccer robots. The experimental results show that objects can be detected and their 3-D information can be obtained accurately, satisfying the real-time requirements of the MSL competition and decreasing the robot's onboard computer's CPU burden. In addition, the proposed algorithm for obstacle detection is not dependent on a specific color.