Performance evaluation of multiple sports player tracking system based on graph optimization
Yuri Nishikawa, Hitoshi Sato, Jun Ozawa · 2017
Multiplayer tracking is one of the essential technologies in team sports analysis, but there are few commercial solutions that consumers can quickly obtain and install. We focused on multiple object tracking method using graph optimization algorithm called K-Shortest Paths (KSP), which performs well with cameras located at a tripod-level height where occasional detection failure may occur. We evaluated end-to-end execution time and accuracy to verify the feasibility of KSP as a multiplayer tracking method. As a result, using a high-performance computing environment equipped with 8 NVIDIA Tesla P100 GPUs, we found that tracking a basketball video of 1 hour is expected to be completed in about 2 1/2 hours. On the other hand, nearly 30 ID switches (replacement of individual trajectory) appeared in 60 seconds. We found that about 75% of the case was either due to the system tracking one person as multiple persons or when people were in contact with each other.