Deep Learning-driven Blind Spot Detection for Forklifts in Industrial Environments

Abstract In modern industrial settings, ensuring the safety of human workers coexisting with heavy machinery like forklifts is paramount. Traditional safety measures often fall short in addressing dynamic hazards, especially in blind spot areas where visibility is limited. This research introduces an advanced blind spot detection system utilizing deep learning techniques specifically designed for industrial environments. The system leverages the YOLOv8 architecture, enhanced through transfer learning, and optimized with model pruning and quantization techniques to achieve high accuracy and low latency, operating at 45 FPS on edge devices with a mean Average Precision (mAP) of 97.3%. The detection models are integrated with a real-time video processing pipeline using the RTSP protocol and OpenCV for spatial awareness. The system incorporates a novel attention mechanism to improve detection accuracy in challenging conditions, such as occlusions and varying lighting. Real-time alerts are provided via an integrated hardware system using ESP8266 microcontrollers and Node-RED for orchestration, ensuring immediate hazard notifications to operators. Extensive experiments in an automotive wheel manufacturing facility validate the system’s effectiveness, demonstrating an improvement of up to 5% in mAP compared to traditional methods. This work contributes to a scalable, efficient, and highly accurate solution for enhancing safety in high-risk industrial environments by addressing critical blind spot hazards in real time.