Computer Vision-Based Collision Prevention in Construction: Bridging Academic Advances and Industrial Requirements

Computer vision (CV) technologies have emerged as promising tools for real-time collision risk management. This paper examines CV-based collision-prevention systems from academic and industrial perspectives with a focus on identifying technical and implementation challenges affecting their deployment on construction sites. A scoping review of 158 publications was conducted, alongside an assessment of commercial CV-based collision-prevention systems currently used in construction. The analysis focused on five key functions: object detection, distance measurement and proximity estimation, tracking, worker re-identification (ReID), and deployment robustness. Academic studies demonstrate substantial progress in detection accuracy, real-time processing, trajectory tracking, and predictive risk assessment through deep learning approaches. However, performance remains sensitive to occlusions, lighting variations, environmental conditions, and site-specific characteristics. The commercial systems assessed in this study, while generally designed for reliability and ease of integration, often rely on simplified zone-based warning logic and provide limited transparency regarding performance validation, tracking capabilities, and distance estimation accuracy. The review highlights a persistent gap between laboratory performance and operational effectiveness on construction sites. The findings point to the need for greater robustness under real-world conditions, validated metric-distance accuracy, improved temporal continuity through tracking and ReID, and standardized evaluation protocols incorporating safety-relevant performance metrics. As AI-enabled systems increasingly support safety-related functions, their development should also align with established machine-safety principles and emerging standards.