Individual Player Action Recognition Using Keypointbased Unweighted Graphical Approach In Sports

Sports involve a variety of interactions between people and objects with little change in posture. Visual classification makes an effort to comprehend and recognize these actions. There are two main processes in the identification of human–object interaction (HOI). The initial step in HOI is localizing the subject and the object. The second element of it involves the identification of complicated relationships between them. Hence, the proposed work focused on detection on the player actions individually by considering their body gesture across the entire video. To attain this, a novel approach is proposed which localizes the joint points with normalized values using KDCnet architecture. Using the plotted points an unweighted graph-based approach is applied to determine the resultant posture labels. At last, the action class is defined by using the body part labels along with posture combination LSTM. The proposed keypoint mapping and action 162recognition (KM-AR) framework provides a meaningful classification result as the graph follows single source traversal along with minimal path coverage across the plotted points. Experiments were conducted using two action datasets such as Sports Video Wild (SVW) and MPII for both keypoint plotting and action detection. The performance of the proposed work is evaluated by comparing with other state-of-the-art methods used for action recognition in the field of sports.