Abstract:Accidents of the rail electric locomotive in the underground are frequently occurring due to the poor coal mine roadway environment and fatigue-limited manual driving. To address this issue, an improved YOLOv4-Tiny framework, called YOLOv4-Tiny-4S, is proposed to achieve multi-object real-time detection for rail electric locomotive. A 4-scale prediction head is designed to enhance the detection ability of small objects based on the 2-scale prediction of YOLOv4-Tiny. Spatial pyramid pooling (SPP) is introduced into the neck network to extend the receptive field and enrich the fused features. In this work, the detection objects concern person, locomotive, stone and lamp in the coal mine roadway, and a new dataset is constructed to validate the proposed approach. The K-means and K-means++ algorithms are applied to re-cluster the samples, and K-means++ has a better performance in this dataset. The ablation studies over YOLOv4-Tiny are conducted to validate the technical improvements in this work. The result comparison over other baselines demonstrates that the proposed approach can detect and recognize the required objects in the underground rail electric locomotive with considerable high performance, achieving 95.35% mean average precision. Specifically, the average precision for small object “stone” reaches 86.69%, yielding 12.38% and 41.66% improvements over YOLOv4-Tiny. The average detection speed of the proposed approach is 58.7 frames per second (FPS), with only 26.3 Mb memory. The proposed multi-object real-time detection method is expected to provide technical support for the unmanned driving of mine electric locomotive.