An Approach to Resource-Efficient Optimization for Real-time Computer Vision-based OHS Monitoring on Resource-constrained Industrial Objects

Abstract

This paper proposes and investigates a methodology for the resource-efficient optimization of real-time AI-based computer vision technologies for occupational health and safety (OHS) monitoring tasks on industrial objects, specifically targeting deployment in resource-constrained systems. The core of the methodology enables the efficient use of CPU-based server devices by introducing two key improvements: a two-stage detection mechanism with dynamic Region of Interest (ROI) selection to significantly reduce the computational load on the primary object detection model, and a static background subtraction algorithm for pre-filtering the video stream and removing non- informative scene areas. A detailed analysis of the performance (in terms of processing speed for real-time capability) and accuracy of various configurations implementing this approach, based on models from the YOLO family, is conducted. A conceptual scheme for integrating the proposed optimization techniques into a typical video analytics pipeline is described, along with a methodology for creating and annotating a specialized dataset. The paper demonstrates that the developed methodology achieves an acceptable balance between violation detection accuracy and video stream processing speed necessary for real-time operation on CPUs, opening prospects for the wider adoption of intelligent safety systems in environments with limited computational resources.