A Decoupled Streaming Architecture for Cross-View Player Association: System Design and Benchmark Protocol.
Abstract
Matching players between camera feeds in sports video is usually done with deep person re-identification (Re-ID) networks such as OSNet, which add a second learned model to the inference path. This report specifies a decoupled alternative for a broadcast camera and a tactical overhead camera. A YOLOv8 detector writes player boxes to disk. A second pass streams the video, links boxes into tracklets with a Kalman filter and IoU (no appearance model), and computes one 8x8x8 HSV histogram per tracklet. Cross-view tracklets are then matched by a temporally gated Hungarian assignment with dummy rows and columns for players visible in only one view. We give the assignment formulation, the memory behavior of each stage, and a benchmark protocol for comparing accuracy, runtime, and memory against an OSNet baseline on edge hardware. No empirical results are reported in this version. Color histograms are expected to separate teams but not teammates in identical kits.