SPL-BEV: Soccer Player Localization and Birds-Eye-View Estimation

Ivar Persson, Håkan Ardö, Mikael Nilsson

Abstract: In this work we present SPL-BEV, a method to localize soccer players on a pitch from a monocular RGB camera. SPL-BEV features a network with few parameters that does not need to make any explicit object detection before localization is made. With SPL-BEV we show increased performance on the Spiideo SoccerNet SynLoc dataset compared to the best provided baseline result. The SPL-BEV system samples features from the U-Net feature space using bi-linear interpolation, guided by camera calibration, to generate features at grid points across multiple planes in a 3D world coordinate system. This forms a voxel feature space, which is then processed into grid cell detections on the ground plane, with final location refinement through x/y correction. The code for SPL-BEV is also published open source on GitHub.

External IDs:doi:10.1007/978-3-032-04968-1_10