Attention-driven multiobject tracking for proactive fall safety management in construction

Falls from height remain a critical safety challenge in the construction industry, where vision-based monitoring systems must balance accuracy, efficiency, and reliability. This study proposes and validates an end-to-end framework for proactive fall prevention that directly addresses this trade-off. First, the framework introduces an efficiency-aware detector, You Only Look Once version 8s for Construction Worker Safety (YOLOv8s-CWS), with targeted architectural enhancements to improve accuracy on small and occluded workers. Second, it establishes a synergistic design where this high-precision detector is paired with the motion-centric ByteTrack algorithm to achieve reliable identity tracking amidst the visual clutter of dynamic construction sites. Finally, it incorporates a validated, building-information-modeling-agnostic method for real-time hazard judgment, enabling safety personnel to define unprotected edges and openings directly on a live video feed. System performance was validated on challenging construction site video sequences. The proposed YOLOv8s-CWS detector achieved a mean average precision of 83.7%, a +1.5% point improvement over its baseline, while operating at 142.7 frames per second. This enhanced detection directly mitigated tracking failures, boosting the system's identity F1 score to 75.32%, for a +7.85 point gain over the baseline configuration. Field deployments confirmed the framework's effectiveness in delivering robust, real-time intrusion alerts. This study provides a computationally efficient and validated solution for proactive safety monitoring, offering a practical tool to mitigate fall-related risks on construction sites.