Automatic Shuttlecock Motion Recognition Using Deep Learning

Abstract

In the field of sports video processing, specifically in the context of motion recognition for shuttlecock match videos, we first propose a method based on attitude estimation to measure the movement extent of the shotting arm, allowing for temporal positioning of shuttlecock shot motions and extraction of corresponding shot-sequences. The shot-sequences is defined in this paper as video segments that exclusively contain the complete shot motion of the mainplayer. During the training phase, we incorporate a lightweight channel-spatial attention mechanism into the Temporal Segment Network (TSN) to classify the extracted shot-sequences into four types, i.e., forehand shot, backhand shot, smash shot, and drop shot. Furthermore, we employ image morphology-based techniques to further distinguish smash shot-sequences as either high clear shots or kill shots based on the shuttlecock's motion trajectory. The extensive experiment results demonstrate the effectiveness of proposed method in accurately positioning and recognizing shuttlecock shot motions.