Learning to Detect Soccer Pass at the Edge: A Comparison Between Edge and Cloud-Based Streaming Machine Learning Approaches

Sports analytics is experiencing substantial growth thanks to the flow of data from wearable tracking devices. Simultaneously, as Edge Computing and the Internet of Things (IoT) gain traction, edge processing models bring computation closer to data origins, cutting down on latency and reliance on cloud services. This paper investigates the effectiveness of Streaming Machine Learning (SML) and Edge Computing in sports analytics, particularly in detecting passes during soccer matches. This research proposed several SML models trained directly on a microcontroller, using the Motus ml proprietary libraries TinySML and TinySML-P, aiming at distinguishing between passing and non-passing actions from data on leg movements collected from sensors embedded in players’ footwear. We compare these algorithms to well-known SML libraries trained on a quad-core/16GB RAM machine hosted in the cloud, evaluating their classification performance and resource usage. Findings show that i) TinySML models executed in the cloud statistically outperform the state-of-the-art ones, and ii) TinySML-P is the only one able to run on a microcontroller, achieving statistically similar results w.r.t. TinySML executed in the cloud. This establishes TinySML-P as a real-time sports analytics solution within resource-constrained edge devices.