CN-116152741-B - Underground personnel detection method based on deep learning algorithm

CN116152741BCN 116152741 BCN116152741 BCN 116152741BCN-116152741-B

Abstract

The invention discloses an underground personnel detection method based on a deep learning algorithm, and relates to the technical field of target detection. The method comprises the following steps of 1) obtaining an underground monitoring video, 2) obtaining an underground personnel monitoring image according to the underground monitoring video obtained in the step 1), carrying out brightness enhancement on the underground personnel monitoring image, and 3) obtaining a detection result of underground personnel based on YOLOX networks. And 2) performing frame cutting on the underground monitoring video to obtain a dim light image, marking personnel and equipment information in the dim light image to obtain an underground personnel monitoring image, and performing brightness enhancement on the image by using a deep learning network based on a U-Net structure. Step 3) YOLOX the network contains a backhaul component, a Neck component, and a DecoupleHead component. According to the invention, on YOLOX frames, an image brightness enhancement algorithm is effectively combined, and the detection capability of a target detection algorithm is further improved by preprocessing the input data for enhancing the dim light image, so that the detection task is accurately and efficiently executed.

Inventors

GE YANG
ZHU JIAHUI
YANG LE
TAO LEI
LIU FENG
WANG HONGWEI

Assignees

太原理工大学

Dates

Publication Date: 20260508
Application Date: 20230223

Claims (4)

1. The underground personnel detection method based on the deep learning algorithm is characterized by comprising the following steps of: 1) Acquiring an underground monitoring video; 2) Obtaining an underground personnel monitoring image according to the underground monitoring video obtained in the step 1), and carrying out brightness enhancement on the underground personnel monitoring image; The method for obtaining the monitoring image of the personnel under the mine comprises the steps of carrying out frame cutting on the monitoring video under the mine to obtain a dim light image, and marking personnel and equipment information in the dim light image to obtain the monitoring image of the personnel under the mine; The method comprises the steps of carrying out brightness enhancement on an underground personnel monitoring image by using a deep learning network based on a U-Net structure to obtain a training set based on an underground monitoring video, wherein the training set comprises a first public data set which comprises an underground personnel image and a label, and a second private data set which comprises the underground personnel image with enhanced brightness and is marked with a label by a professional, and the label comprises a personnel label; The dark light image is enhanced in brightness by a deep learning network based on a U-Net structure through semantic segmentation, an up-sampling layer of the deep learning network is replaced by an deconvolution layer, a convolution layer of the deep learning network uses hole convolution, the deep learning network further comprises a brightness evaluation layer, the brightness evaluation layer groups the features obtained by the convolution of the last layer of the deep learning network, space-time consistency and color consistency image parameters are calculated according to the feature groups, a formula of an image brightness curve evaluation function is used in the brightness evaluation layer, A=A+α(1-A); Wherein A represents a given input, alpha represents a pixel gradient for controlling the exposure level, and the value range of alpha is [ -1,1]; 3) Based on YOLOX networks, detection results of underground personnel are obtained.
2. The method for detecting underground personnel based on the deep learning algorithm of claim 1, wherein in the step 3), the YOLOX network comprises a backup component, a Neck component and a DecoupleHead component; The backup assembly comprises a CBL module, a CSP module and an SPP module which are sequentially connected, the Neck assembly comprises a CBL module, a Cat module and a DeConv module, the DecoupleHead assembly comprises a CBL module and a Decouple module, the backup assembly further comprises a convolution layer, a normalization layer, an activation layer, an up-sampling layer pooling layer and a residual link, and the Neck assembly further comprises an up-sampling layer.
3. The method for detecting underground personnel based on the deep learning algorithm of claim 2, wherein in the step 3), the YOLOX network is trained by using a loss function with edge detection capability, wherein the formula is that, ; Wherein L represents a loss function having an edge detection capability; Representing a preliminary classification loss; representing a preliminary positioning loss; representing a classification loss between the boundary prediction result and the real data; Representing a loss of positioning between the boundary prediction result and the real data, P B representing a predicted boundary classification score, C * representing a real bounding box; Representing the average loss of positive sample number.
4. The method for detecting underground personnel based on the deep learning algorithm of claim 1, wherein when the underground personnel image is detected to contain personnel, whether the personnel is in a dangerous area or dangerous behavior is judged according to information of a video source.

Description

Underground personnel detection method based on deep learning algorithm Technical Field The invention relates to the technical field of target detection, in particular to an underground personnel detection method based on a deep learning algorithm. Background Target detection is a very important technology in deep learning, and since the R-CNN in 2014, the target detection technology based on deep learning is rapidly developed, and is also a core algorithm of an intelligent monitoring system and an important branch of image processing and computer vision disciplines. The target detection technology is divided into two types from the stage, namely a single stage and a double stage, wherein the core idea of the double stage is to propose proposal frames, the approximate position, the size and the probability of being the foreground of the target frame are regressed through a network in the first stage, the position, the size and the category of the target frame are regressed through another network in the second stage, and the core idea of the single stage network is to regress the size, the position and the category of the target in the image through the network directly. The single-stage has the characteristics of high detection speed, high precision and the like, so that the single-stage can be widely applied to actual production environments. YOLOX is an improved version of the well-known single-stage object detection algorithm YOLO, which mainly improves the data enhancement capability of the original YOLOv network, increases the decoupling of the Head stage, and improves the label distribution strategy. With these improved implementations, end-to-end detection can still be performed. Dark image enhancement algorithms are used to enhance useful information in an image, the purpose of which is to improve the visual effect of the image. The conventional image enhancement algorithms based on the neural network comprise an SRIE algorithm, a LIME algorithm, a RetinexNet algorithm, a ENLIGHTENGAN algorithm and the like, and the algorithms are continuously improved along with the development of technology. At present, an underground personnel detection algorithm takes real-time video image data and an intelligent image recognition technology of an underground coal mine camera as cores, and mainly utilizes camera equipment with better performance to try to solve the problem of poor underground visibility, but the cost of the solution is higher. While the detection of underground personnel can also be considered as a specific application under target detection, the algorithm for achieving brightness enhancement and personnel detection end-to-end with deep learning has not been truly applied to the detection of underground personnel, and therefore the method will be applicable to underground personnel. Disclosure of Invention The invention provides a submerged personnel detection method based on a deep learning algorithm, which aims to solve the problem that the deep learning algorithm is not applied to submerged personnel detection in the existing target detection technology. The invention is realized by the following technical scheme that the submerged personnel detection method based on the deep learning algorithm comprises the following steps: 1) Acquiring an underground monitoring video; 2) Obtaining an underground personnel monitoring image according to the underground monitoring video obtained in the step 1), and carrying out brightness enhancement on the underground personnel monitoring image; ① The method for obtaining the underground personnel monitoring image comprises the steps of carrying out frame cutting on the underground monitoring video to obtain a dim light image, and marking personnel and equipment information in the dim light image to obtain the underground personnel monitoring image. ② The method comprises the steps of carrying out brightness enhancement on an underground personnel monitoring image by using a deep learning network based on a U-Net structure, processing image data acquired in a severe underground environment by using a dim light image enhancement algorithm, improving image quality and definition, detecting by using an auxiliary target detection algorithm, calculating and analyzing personnel information in a current input image, demarcating a dangerous area, warning personnel in the dangerous area, obtaining a training set based on an underground monitoring video after brightness enhancement, wherein the training set comprises a first public data set comprising an underground personnel image and a label, the training set further comprises a second private data set comprising the underground personnel image subjected to brightness enhancement, and calibrating the label by a professional, and the label comprises a personnel label. ③ The dark light image is enhanced in brightness by a deep learning network based on a U-Net structure through semantic segmentation, an up-sampling la