Tennis Patterns Recognition Based on a Novel Tennis Dataset – 3DTennisDS

Many scientific studies on tennis stroke recognition are based on datasets created for the purpose of research using video or motion capture techniques. The importance of such datasets has been increasing due to the athlete performance evaluation needs. The primary aim of this paper is to present a state-of-the-art 3DTennisDS storing four tennis strokes: forehand, backhand, volley forehand and volley backhand. The moves were registered using the Vicon optical motion capture and contain a 39-marker player and a 7-marker tennis racket models. The potential and quality of this unique dataset has been verified using Spatial-Temporal Graph Neural Networks, because this type of network topology matches to the human body structure. The presented 3DTennisDS has been compared with two well-known datasets: the THETIS and the Tennis-Mocap. They contain tennis movements in a form of motion capture data, registered using markerless and marker-based systems. The classification of tennis strokes has been performed to verify how various types of data acquisition (marker-based and marker-less ones) as well as the structure of the data affect the accuracy of human action recognition. In this study ONI files from THETIS, bvh from Tennis-Mocap and c3d data from 3DTennisDS were considered. Moreover, the impact of input data fuzzification was examined. The obtained results showed that the classification using 3DTennisDS achieved the best results, both for fuzzy and non-fuzzy inputs. These outcomes indicate that the way of capturing data, its preparation and structure have great influence on classification accuracy. The developed 3DTennisDS has a great potential in further motion capture analysis.