Feature Fusion based Efficient Convolution Network for Real-time Table Tennis Ball Detection

Abstract

Ball detection is an important technique for the athlete training and technical analysis of the sport like table tennis, especially for the ball trajectory real-time analysis for the table tennis robot or real-time technical guidance system. How to detect the flying ball accurately and timely are two main challenges in this area. Since the table tennis ball is relatively small compared with the playing court, it belongs to the category of small object detection in large scenes. In this study, an efficient one-stage based detection algorithm structure is proposed to speed up the detection and achieves promising accuracy. Specifically, a network pruning strategy is applied to accelerate the detection. Furthermore, in order to alleviate the performance drop caused by network pruning, a novel context fusion component is introduced to the network which is a specified design for small object detection. The experimental results show the proposed feature fusion based efficient convolution network achieves real-time table tennis ball detection of 3.4 ms on the Nvidia 1050Ti hardware with 0.971 (mAP) accuracy.