Deep HM-SORT: Enhancing Multi-Object Tracking in Sports with Deep Features, Harmonic Mean, and Expansion IOU

This paper introduces Deep HM-SORT, a novel online multi-object trackingalgorithm specifically designed to enhance the tracking of athletes in sportsscenarios. Traditional multi-object tracking methods often struggle with sportsenvironments due to the similar appearances of players, irregular andunpredictable movements, and significant camera motion. Deep HM-SORT addressesthese challenges by integrating deep features, harmonic mean, and ExpansionIOU. By leveraging the harmonic mean, our method effectively balancesappearance and motion cues, significantly reducing ID-swaps. Additionally, ourapproach retains all tracklets indefinitely, improving the re-identification ofplayers who leave and re-enter the frame. Experimental results demonstrate thatDeep HM-SORT achieves state-of-the-art performance on two large-scale publicbenchmarks, SportsMOT and SoccerNet Tracking Challenge 2023. Specifically, ourmethod achieves 80.1 HOTA on the SportsMOT dataset and 85.4 HOTA on theSoccerNet-Tracking dataset, outperforming existing trackers in key metrics suchas HOTA, IDF1, AssA, and MOTA. This robust solution provides enhanced accuracyand reliability for automated sports analytics, offering significantimprovements over previous methods without introducing additional computationalcost.