Abstract
Construction site safety remains a critical global challenge, demanding urgent attention. Existing surveillance systems struggle to balance multi-object detection accuracy, real-time efficiency, and environmental robustness under strict edge constraints. This paper presents MMDet-Edge, an edge-optimized unified detection framework that addresses these competing demands via three synergistic innovations. First, an adaptive feature fusion architecture with a learnable spatial-channel attention mechanism resolves cross-scale conflicts, boosting small-object average precision (AP) by 9.3%. Second, a hardware-conscious neural architecture search (HC-NAS) strategy co-optimizes sparsity patterns and quantization sensitivity, achieving a state-of-the-art performance of 89.4% mAP@0.5 at only 1.8 W power consumption-surpassing contemporary edge detectors by 6.3% mAP under equivalent power budgets. Third, by incorporating OSHA fatality statistics into a novel risk-weighted evaluation paradigm, we reduce high-consequence false negatives by 34%. Comprehensive evaluations on a purpose-built benchmark and cross-dataset tests demonstrate MMDet-Edge's superiority. It outperforms a wide range of state-of-the-art models. Validated across three active construction sites, the system enables real-time detection of five safety-critical targets (personnel, helmets, flames, smoke, vests) under extreme conditions, including >60% occlusion and >100 lux illumination variance. Our field deployments demonstrated a 22% reduction in safety incidents compared to conventional systems, establishing a new architectural paradigm for safety-critical edge AI through principled hardware-algorithm co-design.