Measurement of elite Taekwondo athletes using convolutional neural networks in competition environments

Elite Taekwondo performance analysis requires valid and scalable ways to measure athlete positions and interpersonal distances directly from competition video. This research study explores whether convolutional neural networks (CNNs) can achieve such measurements in authentic, single-camera environments that include occlusions, referees, and changing viewpoints. To answer this, first this study examined what manual annotation can achieve: expert annotators provided the benchmark for reliability, and elite coaches classified distance categories to assess their qualitative agreement and how these related to real-world ranges. Agreement between coaches was only moderate, showing broad overlap in perceived distance bands, suggesting that fixed thresholds are unsuitable without coach-specific calibration. Building from this manual baseline, the extent to which automated systems could replicate or improve upon these capabilities was analysed. Manual annotation demonstrated high reliability in favourable viewing conditions but deteriorated with greater occlusion and shallower camera angles. These findings defined the standard against which automation was assessed. A custom YOLO detector and subsequent pose estimation pipeline were developed to automate athlete localisation and distance estimation. Detection provided a robust basis for identifying athletes, while pose estimation refined positional accuracy by capturing body landmarks. Combining the two within a hybrid framework reduced identity swaps and stabilised distance estimates across diverse scenarios. Overhead or elevated audience views supported the most consistent results, whereas broadcast footage with heavy occlusion remained challenging. Overall, 2D CNN-based measurement in elite Taekwondo proved it was able to accurately measure when supported by careful calibration and a scenario-aware processing pipeline. Automated systems can objectively extract metrics with the equivalent accuracy of manual reliability under optimal viewpoints, though precise distance inference remains difficult in complex visual conditions. Practically, these tools should report continuous distance metrics and, where categorical outputs are desired, apply coach-calibrated mappings to ensure interpretive validity. This approach aligns automated performance analysis with human perception while accommodating the visual and tactical complexities of real competition.