Lightweight real-time object detection for coal mine underground unmanned vehicles based on improved YOLOv8

Accurate object detection provides reliable environmental information for unmanned vehicles, serving as a critical safeguard to achieve safe and efficient operation of autonomous underground mining vehicles. However, traditional detection techniques face challenges such as low accuracy and poor real-time performance in complex underground coal mine environments characterized by insufficient lighting, coal dust interference, and dense distribution of small targets. This paper proposes an enhanced YOLOv8n-based algorithm incorporating three optimization strategies: 1) Replacing standard convolutional layers with depthwise separable convolutions (DSC) in the backbone network to significantly reduce parameters while maintaining feature extraction capabilities; 2) Developing a novel SPPF-LSKA module that integrates lightweight spatial pyramid pooling with attention mechanisms, achieving balanced multi-scale feature fusion and computational efficiency through ReLU activation optimization; 3) Introducing the Extended Intersection over Union (EIoU) loss function to improve bounding box regression accuracy under occlusion conditions. Experimental evaluations on a self-built coal mine dataset demonstrate substantial improvements over the original YOLOv8n: 5.0% higher precision, 5.9% greater recall, and 5.8% enhanced [email protected], while maintaining real-time detection performance at 108.8 FPS. These results validate the proposed method's effectiveness for industrial inspection tasks in complex environments. This method effectively addresses missed and false detections in complex underground mining scenarios, providing a high-precision, lightweight object detection solution to facilitate the standardised deployment of unmanned vehicles in coal mine environments.