RGBD image based human detection for electromechanical equipment in underground coal mine

Human detection within the operating area of electromechanical equipment is essential to ensure safe production and avoid accidents in the underground coal mine. Low light intensity and uneven light distribution in its environment surrenders the traditional color image based methods for human detection. In this paper, we focus on accurate detection of human in the operating area of electromechanical mining equipment using RGBD image. A novel network framework for miner detection based on YOLOv3 is proposed to fuse color image and depth image with enhanced attention mechanism. In the Pre-Backbone, feature extraction of both Depth and RGB branches are developed as the preliminary feature extractor with convolutional layer and residual block. Then the Convolutional Block Attention Module (CBAM) is improved to select and fuse RGB and Depth features by defining channel weights. Finally, the features are further inputted to Post-Backbone and used for multi-scale prediction in Head. The experiments demonstrate the superiority of the proposed method over some classical methods on miner detection with different light intensities and distributions.