Unmanned driving in underground mining area is an important component of the intelligent mining transportation system and a key link in the construction of intelligent mine.The real-time intelligent perception in the complex environment of mine roadway can provide important information for the decision-making system and driving system of unmanned vehicles,which is a necessary condition for the safe driving of unmanned vehicles in underground mining areas.Compared with laser radar and other perception technologies,environment perception based on machine vision is gradually becoming a mainstream technology for driverless obstacle detection due to its advantages of low cost,easy data processing and abundant target information acquisition.This topic is mainly based on the fusion method of deep learning infrared obstacle detection and binocular vision ranging to study the obstacle target detection of mine roadway in dim light.The main contents include:(1)Unmanned vehicle forward obstacle image data preprocessing and dataset construction.In this paper,the datasets are all collected from the field,and the falling rocks and pedestrian main road obstacles are added in the mine tunnel.In order to solve the problems of low clarity of some images in infrared data,low image contrast,unclear obstacle features and small number of training datasets,the mine tunnel obstacle datasets are processed by some basic image pre-processing methods and image augmentation techniques to lay the foundation for subsequent network training.(2)Research on infrared obstacle detection method of mine roadway based on machine vision.The classic YOLO series algorithm in target detection is selected as the basic detection algorithm.In view of the fact that its original Anchor frame is not applicable to the data set in this paper,optimized K-means++ is used to re-cluster the anchor frame of obstacles in this paper,and an Anchor Box suitable for obstacles in mine roadway is constructed.Deep separable convolution is used in the backbone of the network,which can significantly reduce the number of parameters and computation,and improve the reasoning speed of the network.In the feature fusion stage,a RCR module is proposed based on the combination of residual module and dual-channel attention mechanism.This module can solve the problem of model degradation caused by too deep network,and assign different weights to different regions of the image so that the network pays more attention to important parts,and improves the detection accuracy of infrared image obstacles with little texture.(3)Research on the method of determining obstacles in front of unmanned vehicles in mine tunnel based on infrared binocular vision.The internal and external parameters of the camera are obtained by calibrating the binocular camera.The binocular images are calibrated and binocular images are stereo matched to generate parallax maps to obtain the image coordinate system information of the pixels in the images,i.e.,the measured obstacle distances.The distance measured by using laser rangefinder is compared and analyzed with the distance measured by the binocular distance measurement algorithm in this paper,and it can be seen that the error of the distance measurement algorithm in this paper is small.Finally,the previous obstacle detection algorithm and the ranging algorithm are fused to achieve the accurate measurement of the obstacle category and distance information in front of the mine roadway traveler.Through the analysis of the experimental results,it can be seen that the mine roadway obstacle detection method based on infrared binocular vision proposed in this paper has a good detection effect on mine roadway obstacles.Meanwhile,in terms of the measurement of roadway obstacle distance,its error range is also below the centimeter level,which can meet the accuracy and speed requirements of obstacle detection and ranging when the underground mine unmanned vehicle is moving. |