MS-3D-CNN with quantum-inspired chaotic hybrid adaptive wolf-particle swarm optimization for tactical pattern classification in basketball from player tracking data

Introduction AI sport analysis methods have been considerably improved by developments in knowledge of tactically information and its codification. Nevertheless, methods for capturing multi-levels spatio-temporal information for complex team-sport behaviors are still limited. In opposed to motion recognition, which concentrates on movements of individual athlete, tactically pattern classification refers to the recognition of coordinated group actions based mostly on the athlete space-time trajectories. Methods An efficient Q16 framework based on MS-3D-CNN and Chaotic Opposition-based Adaptive Wolf-Particle Swarm Optimization (QIC-HAWPSO) for action recognition is proposed in this paper. Parallel multi-scale 3D kernels (three different scales of 3, 5, and 7 frames which correspond to 0.12, 0.20, and 0.28 seconds) are used to learn quick individual actions, complex actions, as well as slow formations. QIC-HAWPSO has chaos opposition-based learning, quantum tunneling adaptive mechanism, and population diversity awareness. Results The framework demonstrated an accuracy of 94.3% on the Basketball-51 dataset (10,311 clips for 8 types of tactical action), leading the current 7 state-of-the-art approaches, such as Game Sense (92.8%), ACA-Net (89.3%), and HAWPSO-3D-CNN (90.1%). Moreover, the inference latency can be evaluated at 34.2 ms clip-1 based on precomputed tracking-grid input and a memory size of 2.4 GB. Discussion The classifier will be useful in online coaching, broadcasting, and player analysis, depending on upstream tracking data. It consists of a multi-scale MS-3D-CNN framework, a superior QIC-HAWPSO optimizer, and experimental results that demonstrate the importance of the approach over baseline strategies.