SquashSight: A Single-Camera Deep Learning Pipeline for Automated Squash Match Analysis and Coaching Decision Support

Squash is among the least analytically instrumented racket sports, despite consistently producing world-class athletes and gaining global recognition through its inclusion in the 2028 Olympic program. Automated match analysis is complicated by the ball’s small size and high speed, two ball color variants, frequent player occlusions on a confined court, and the near-total absence of publicly annotated datasets. Professional-grade systems rely on expensive multi-camera installations that are inaccessible to most clubs and academies. We present SquashSight, a modular single-camera pipeline that converts standard match footage into structured coaching analytics without specialized hardware. The system is organized as a nine-stage pipeline in which each stage enriches a single per-frame table, and it integrates homography-based court calibration, YOLOv8-Pose player detection with ResNet50 re-identification, a dual-color adaptation of TrackNet for ball tracking, recurrent classifiers for rally segmentation and stroke recognition, trajectory-based signal processing for hit detection, and rule-based shot classification in real-world court coordinates. A natural-language coaching assistant, constrained to query the extracted data through deterministic analytics tools, provides a grounded decision-support layer on top of the pipeline. Evaluated on five annotated squash match videos, the pipeline reaches an F1-score of 0.973 for player tracking (MOTA 0.943), high-precision ball tracking (precision 0.91 to 0.94, located within 10 px of ground truth 94 to 96 percent of the time, though recall is limited to 0.45 to 0.48 on fast-motion frames), and 1.000 rally segment detection with 0.853 boundary accuracy, while processing at roughly twice real-time speed on a single consumer GPU. As a proof-of-concept result on a single held-out video (n=122), the stroke classifier separated forehands from backhands with no misclassifications. The results indicate that targeted domain adaptation of established deep learning models can support accessible, interpretable coaching decision support for an under-instrumented sport.