基于计算机视觉技术的矿井电机车无人驾驶障碍物检测技术

Obstacle Detection Technology of Mine Electric Locomotive Driverless Based on Computer Vision Technology

针对传统计算机视觉方法难以实现障碍物实时检测和定位的问题,提出传统计算机视觉技术与深度学习目标检测算法YOLOv3相结合的障碍物智能检测方法。首先,采集电机车行驶区域(即有效检测区域)存在的障碍物数据并制作VOC格式数据集,使用YOLOv3训练数据集,得到障碍物检测模型;然后,采用传统计算机视觉技术定位到轨道,使用"3邻域"搜索法获得轨道线坐标值,根据距离信息向轨道外侧扩展一定距离,提取有效检测区域,同时网格化图片,将障碍物的坐标换算为实际距离;最后,使用障碍物检测模型对有效检测区域进行检测。试验结果表明:该方法可以识别行驶区域内多种特征差异很大的目标物体,如电机车、人和大块落石等;该方法每秒可以处理6帧图片,现场采集的实际数据测试平均精确率达到93.2%。

The mining of mineral resources is getting deeper and deeper,the working environment is bad,the employees are aging seriously,and the human cost is rising,which brings great challenges to the development of mining industry.Intelligent mine operation has become an inevitable trend.As a part of mine intellectualization,unmanned transportation system is very important for mine,which means the improvement of safety production efficiency can achieve zero injury and zero time loss.Considering the development demand of driverless electric vehicle and the traditional computer vision method,it is difficult to realize the real-time detection and location of obstacles.An intelligent obstacle detection method based on the combination of traditional computer vision technology and deep learning target detection algorithm YOLOv3 was proposed.First of all,the video data of obstacles in the driving area of electric locomotive(this paper calls it effective detection area)were collected,using an image annotation tool namely labelimg to make VOC data set,using YOLOv3 to train data set,according to the feedback results,adjust the parameters continuously to obtain the relative optimal parameters,and finally get the obstacle detection model.Then use the traditional computer vision technology to locate the track by edge,texture and other information,using the"3 neighborhood"search method to get the track line coordinate value of left and right track lines,and expand a certain distance to the outside of the track according to the distance information,extracting the effective detection area,thus reducing the computational complexity of the later obstacle detection,at the same time,gridding images,converting the coordinates of obstacles to the actual distance.Finally,using the obstacle detection model to detect the effective detection area and respond to the detection results.Experimental results show that the method can identify many objects with different characteristics in the driving area,such as electric locomotive,people,large falling rocks,etc.It can process 6 pictures per second,and the average accuracy of the actual data collected in the field reaches93.2%,it has good performance in real-time and accuracy,and has a good effect in the underground mine scene.