CARE-YOLOPose: a structure-aware framework for robust dump truck bed keypoint and edge detection under adverse industrial conditions

Structural perception of dump truck beds is vital for autonomous industrial operations yet remains vulnerable to physical visual degradation and geometric anisotropy. To address these challenges, we propose CARE-YOLOPose (Context-Aware Regression Enhanced YOLOPose), a structure-aware framework tailored for robust keypoint and edge detection. The architecture integrates MH-REAM to suppress high-frequency environmental noise and MH-CCAR to refine slender, coordinate-sensitive structural boundaries. We also introduce DTB-CornerSet, a physically grounded benchmark with 5,000 images spanning seven adverse conditions. Crucially, we propose the Line Similarity (LS) metric. Unlike soft additive metrics, LS employs a multiplicative “Logical AND” mechanism that acts as a strict veto for system-level geometric validity, forcing the model to simultaneously satisfy endpoint, midpoint, and angular constraints. Extensive experiments demonstrate that CARE-YOLOPose achieves an Average LS of 0.955 (Rear Edge LS 0.974), significantly outperforming recent lightweight Transformers in structural fidelity while minimizing computational latency. Quantitative stress tests reveal that the system maintains structural integrity up to a 60% occlusion ratio, defining a critical perceptual breakdown threshold for industrial safety. With a real-time inference speed of 39.1 FPS on a standard GPU and a safety-compliant 13.4 FPS (74.6 ms latency) on embedded Jetson Orin NX edge devices, the framework provides a viable, high-precision observation layer for closed-loop industrial automation.