YOLOv5 Based Pedestrian Safety Detection in Underground Coal Mines
Yang Zhang, Yimin Zhou
Abstract
Safety detection is important for preventing accidents occurence in underground coal mines (UCM). However, the safety detection in UCM could be seriously interfered by complex environmental factors, i.e., dim light and dense dust. In this paper, we construct a simulated dataset for pedestrian safety detection in UCM (SDUCM-dataset), which can not only simulate real coal mine scenes, but also satisfy common safety detection tasks requirement, so as to solve the problem of lack of public coal mine environment dataset. Moreover, a novel YOLOv5 (You Only Look Once version 5) based detector named YOLo-UCM (YOLOv5 based safety detector in UCM) is well designed to detect whether there are potential safety hazards which can achieve higher efficiency than most one-stage object detectors and maintain comparable accuracy compared to the two-stage object detectors. Vision Transformer, Merge Non-Maximum Suppression and Meta-AconC (Activate or not version C) are further introduced to improve the model performance with higher detection accuracy but without significant loss of the detection speed. Experiments are performed to verify the efficiency of the proposed method for pedestrian safety detection in UCM, which demonstrate that the SDUCM-dataset can support the training of the YOLo-UCM, achieving delicate balance between the accuracy and speed in UCM environments.