To improve the accuracy of miners' helmet-wearing recognition in coal mine underground operations，a helmet-wearing recognition algorithm was proposed based on human posture analysis and machine vision system optimization methods. First，Single Shot MultiBox Detector (SSD) was used as the basic model of multi-target identification. The squeezed neural network (SqueezeNet) was used to reduce the model parameters to develop an efficient recognition model，which improved the recognition accuracy of the miners' helmet and maintained the balance between the recognition accuracy and the calculation speed. Then，a multi-person posture estimation algorithm was used to locate the joint points of each miner and determine the miners' complex behavioral status. Finally，the upper limb nodes of the target were extracted based on fusion model of multi-target recognition and multi-person posture estimation，and then the helmet-wearing condition was determined by the spatial topological relationship between the upper limb nodes and the helmet frame. Moreover，3 000 pieces of image data were selected to validate the proposed method's performance. The results indicated that the machine vision system can optimize the hardware and software configuration to improve the system's performance. Furthermore，the identification accuracy reached up to 91.1%，which was much better than that of the latest helmet-wearing recognition algorithm. Therefore，the proposed system in this study can meet the requirements of helmet-wearing recognition accuracy for underground coal miners.