GPU Parallel Computing for Detection and Classification Object in Robot Soccer ERSOW

Abstract

ERSOW robot soccer that participated in the Indonesian Wheeled Robot Soccer Contest, has many abilities such as object detection and classification, control and navigation system, self-localization and mapping, and also real-time communication between each other. This research focusing on object detection and classification on the robots as one of important processes to provide main data sources for all further actions. Therefore, this process takes longer computation time to detect and classify multiple objects, to improve detection and classification performance speed without sacrifice the accuracy value, we proposed GPU parallel computing on offline training phase and online inference phase. In the offline training phase, the neural network model can be trained in parallel processes using selected GPU hardware. As a result of training, we can transfer learning the model knowledge to another host. The experiments on the NVIDIA Jetson AGX Xavier board show that the custom model as the result of the offline training phase achieves more than 30 fps and pre-trained model SSD-MobileNet-v2 achieve more than 99 fps.