Basketball detection and trajectory prediction using IoT for assisting physical training

Abstract

With the development of the Internet of Things (IoTs) and 5G technologies, more and more smart applications are emerging. This paper designs an IoTs-based college basketball teaching system which can automatically detect basketball and predict its trajectory for auxiliary teaching. The difficulties include low-latency video processing and a smart algorithm for automatic basketball detection and its trajectory prediction. For the former issue, the basketball videos are collected using a 5G camera and transmitted to the Jetson TX2 platform through a 5G network. For the latter issue, an end-to-end deep learning framework is proposed and deployed on the Jetson TX2 platform. First, a pre-trained YOLOv5 is used to obtain high-confidence candidate regions; then, the local dependencies are disclosed using a spatial graph convolutional layer; lastly, a multi-head self-attention (MSA) mechanism is used to improve the modeling of long-distance dependencies. The proposed system is evaluated on a self-built basketball dataset and the results show its effectiveness for basketball detection and trajectory prediction.