| Coal is the main energy of China,improving the safety production level of coal resources has always been the focus of national attention,underground personnel positioning is an important part of safe production,in the coal mine underground environment to achieve real-time and high-precision positioning of coal miners to protect the personal safety of coal miners,improve safety production efficiency is of great significance.Compared with other wireless positioning technologies,binocular visual positioning has advantages in positioning accuracy.At present,the research on binocular vision mainly focuses on automatic driving,three-dimensional reconstruction and other aspects,and lacks research on the positioning of underground personnel.Affected by the special underground environment,binocular visual positioning is currently facing two problems:(1)The underground miner’s uniform is similar to the color of the environment,and there is interference such as dust,and the concealment of personnel is strong,and the detection of underground personnel is more difficult.(2)The dark light environment underground weakens the reflection ability of miners’ clothing,and the work clothes show weak texture and complex texture state,resulting in the lack of features required for stereo matching,thus affecting the positioning accuracy.This thesis conducts research on the above problems,and the main contents are as follows:(1)Aiming at the problem of personnel leakage detection caused by special underground environment,this thesis proposes a target detection method based on image adaptive enhancement.Using the atmospheric scattering model,the image imaging process and noise characteristics under downhole dust and dark light interference were analyzed,and then a filter was designed to remove dust and improve brightness.A small convolutional neural network is used to optimize the image filter parameters,and the network model is trained by mixed training of public dataset and homemade dataset.The experimental results show that the accuracy of the proposed method is 6.5% higher than that of YOLOv5,which verifies the effectiveness of the proposed method in improving the detection accuracy of underground personnel.(2)Aiming at the problem that the weak texture and complex texture of miners’ uniforms in the low-light environment of underground are caused by false matching and affect the positioning accuracy,this thesis proposes a Disp-Net C stereo matching method based on detail enhancement.On the basis of retaining the computing speed advantage of Disp-Net C network,the original twin dual-link structure is improved.The hybrid attention mechanism is used to strengthen the model’s ability to extract detailed features such as space and channel,which solves the problem of single feature extraction from the original network convolutional blocks.A multi-scale image feature extraction network is proposed,which extracts the information of different scales of the image and fuses it,which strengthens the characterization ability of image features and alleviates the problem of mismatching of complex textures and weak texture regions.Train network models using a mix of public datasets and homemade datasets.The results show that the average error of the proposed positioning method is 8.9cm in the range of 50lx-100 lx,and the positioning accuracy is improved by 5.3% compared with the reference method.(3)Due to the large number of redundant calculations in the stereo matching of the detection frame area,resulting in low positioning speed,this thesis uses the fusion positioning method to solve this problem.By reducing the matching pixels,redundant calculation of parallax values in the image background area and the person edge area is avoided,and the positioning speed of binocular vision is improved.Finally,the positioning method is tested by experiment.The results show that the positioning method proposed in this thesis and the matching positioning method of the detection frame improve the positioning speed by 45%. |
