Real-time detection algorithm of underground human body based on lightweight parameters

The existing underground personnel target detection methods cannot achieve the real-time detection results due to the deep network and huge calculation amount, a real-time detection algorithm of underground human body based on lightweight parameters is proposed. The method uses the depthwise separable convolution module and the inverted residual module to construct a lightweight characteristic extraction network. Through the depth separable convolution compressing parameter amount and calculation, the operation speed of the characteristic extraction network is improved. The inverted residual structure extracts enough information through a higher dimensional tensor to ensure the accuracy of the characteristic extraction network. Combining the lightweight characteristic extraction network and the SSD multi-scale detection method, an underground human body real-time detection model is established. The model adds traditional convolutional layers to 27 layers to perform convolution operations on the basic structure of the lightweight inverted residual characteristic extraction network. 6-layer characteristic maps are extracted for multi-scale prediction. The test results show that the size of the model is 18 Mbyte, the frame rate is about 35 frames/s, and the performance is better than the commonly used VGG16+Faster R-CNN model and VGG16+ multi-scale detection model. In order to meet the needs of target detection of specific underground environments, a semi-automatic annotation method for human body data based on Faster R-CNN is designed, which can reduce manual workload significantly and improve the accuracy of underground human body detection. The color information of miners' clothing is used for secondary screening of the detection result frame to eliminate the false detection frames that detecting the background as human bodies. The test results show that the algorithm realizes real-time positioning detection and frame selection of mine working face personnel with an accuracy of 92.86% and a recall rate of 98.11%. The algorithm solves the problem of missing and false detection of underground personnel effectively.