Offensive Tactics Recognition in Broadcast Basketball Videos Based on 2D Camera View Player Heatmaps

Tactical analysis is essential in the domain of sports science. Tactical analysis assists teams in reviewing their offensive and defensive execution performance as well as understanding their opponents' tactics in order to generate effective counter strategies. This study focuses on basketball offensive tactics recognition. Most of the current tactics recognition methods are based on top-view trajectory information and rely heavily on accurate sports field camera calibration and player tracking technology to extract trajectory features and learn spatiotemporal correlation models. However, obtaining accurate camera calibration parameters and player tracking data is a challenging task, especially when there are situations such as camera movement, player occlusion, and similar players’ jerseys. The accuracy of the tactic’s identification algorithm is greatly affected. This study attempts to improve the previous tactical analysis algorithm process by omitting the steps of camera calibration and player tracking; directly detecting the position of the offensive players at each time step from the RGB images from the broadcast videos; and mapping the position distribution map of the unsorted offensive players in the camera-view as the tactical analysis model's input. Our proposed method is a recurrent convolutional neural network with coordinate embedding to directly identify the tactics and is combined with the player trajectory reconstruction module, which will aid the model in acquiring a better latent code to represent the tactics. On the validation dataset, for both supervised and unsupervised settings, our proposed method’s performance is comparable to the current tactics classification methods that rely on perfect top-view trajectories as input for the model.