Speaker
Description
Mining operations remain vulnerable to collisions between workers and heavy equipment, particularly in blind-spot areas where conventional personal protective equipment (PPE) provides only passive protection. This paper presents an intelligent safety vest that actively detects hazardous objects and provides real-time warnings to improve worker safety. The proposed system combines a Real-Time Detection Transformer (RT-DETR) for object detection with a Convolutional Long Short-Term Memory (ConvLSTM) network for motion analysis, enabling the system to recognize both the presence and movement of hazardous objects such as trucks, excavators, rocks, and pillars. To enable deployment on a Raspberry Pi 5 edge device, the RT-DETR model is quantized from FP32 to INT8 precision, reducing computational requirements while maintaining reliable detection performance. When a potential hazard is detected, an ESP32 microcontroller activates directional haptic feedback through Taptic Engine actuators, allowing workers to perceive the direction of approaching hazards without relying on visual attention. Experimental results show that the quantized RT-DETR achieves a mean Average Precision (mAP@0.50:0.95) of 0.760, while the ConvLSTM model attains an F1-score of 0.73 and a recall of 0.84 for detecting approaching hazards. In end-to-end system evaluation involving 1,490 test samples, the proposed smart vest achieved an overall accuracy of 66.24% and a recall of 94.35%, demonstrating a strong capability to minimize missed detections, which is critical for safety-critical applications. These results indicate that the proposed wearable edge-AI system provides a practical and proactive solution for enhancing occupational safety in dynamic mining environments.
