Method for real-time boxing punch recognition based on captured motion data

In this paper, we propose a method for real-time recognition of ten typical types of boxing punches from continuously captured boxing sequences. Punches are firstly segmented from continuous boxing sequences by checking the distance between hands and their corresponding shoulders, which is a reliable feature for punch detection and segmentation in continuous boxing sequences. Punches from different subjects, which are of variable length are normalized and uniformly represented with vectors. Then a Two-Step-LDA based punch type recognizer is applied to the unified punches. The type of a specific punch is determined according to a nearest neighbor rule in the LDA feature space. The proposed approach works efficiently and the recognition accuracy is about 98%. We have implemented a real-time punch recognition prototype system, in which the real-time boxing motion is provided by a 7-camera Eagle motion capture system.