Key technologies for identifying abnormal behaviors in fully mechanized mining faces based on panoramic visualization

Abstract

A lightweight human pose recognition model based on OpenPose improvement is proposed to address the difficulty of intelligent supervision of abnormal behaviors of coal mine workers. The backbone network of original model is replaced with MobileNet-v3, and the original 7×7 convolutional kernels is replaced with a smaller convolutional kernel to improve model detection speed and reduce resource consumption. In addition, due to the limited shooting range of the working face camera, it is difficult to meet the needs of panoramic visualization of the working face. This article adopts the SIFT image stitching algorithm to fuse multiple monitoring perspectives and achieve real-time synchronization of full scene monitoring of the working face. After algorithm experiments, the improved human posture recognition model has an accuracy of 81.5%, occupies 42% of the memory of the original model, and the detection speed is 1.59 times that of the original model. The improved human posture recognition model not only reduces the memory consumption, but also greatly improves the speed of model detection and maintains a high accuracy. Combining human posture recognition and image stitching algorithm can realize real-time monitoring of workers’ behaviors.

Keywords

• video monitoring of coal mine working face
• human pose estimation
• image stitching
• openpose
• sift features