Real-Time Human Pose Estimation Based on Selective Knowledge Distillation and Robot Occlusion Detection in Football Stadiums

Abstract

In dynamic, occlusion-prone environments like football stadiums, reliable human pose estimation is essential for mobile robots, where conventional systems often fail due to partial visibility and rapid motion. This paper presents a lightweight pose estimation framework for real-time operation on resource-constrained edge platforms and the X5 robot. It employs selective knowledge distillation to transfer occlusion-robust and motion-aware features from a pre-trained teacher model to a compact student model, preserving efficiency while enhancing reliability. Three components are integrated by synthetic occlusion pattern embeddings for visibility, temporal motion cue extraction for movements and cross-modal attention alignment to focus on visible body regions. Validated on the D-Robotics mono2d_body_detection benchmark, the system achieves accuracy improvements under heavy occlusion while maintaining ≥30 FPS on the target platform. Experimental results confirm the framework balances high accuracy with low complexity, which makes it suitable for reliable deployment in football stadiums.