基于Jetson TX2的矿用巡检机器人行人检测算法的研究与实现

【作者】 李斌；

【导师】 李现国； 乔杰；

【作者基本信息】 天津工业大学 ， 集成电路工程（专业学位）， 2019， 硕士

【摘要】 矿用巡检机器人是煤矿安全生产领域中的重要辅助设备,对提高巡检效率、保障井下安全生产具有重要意义。目前矿用巡检机器人局限于周围环境参数及设备状态信息的采集,不具备对井下工作人员进行实时巡检的功能。同时由于受井下环境恶劣、光照不足、图像背景混杂、图像采集质量差等不利因素的影响,传统行人检测方法不能直接应用于矿用巡检机器人中,也难以实现对巡检沿线工作人员危险行为(比如靠近带式输送机头部、尾部、拉紧装置等易挤夹部位,跨越输送带等)的实时检测和报警,以及工作人员的在岗检查功能。本课题的目的是研究并提出一种基于深度学习的井下行人检测算法,并在嵌入式平台Jetson TX2中部署,实现矿用巡检机器人对井下工作人员的自动检测。首先提出了一种基于Jetson TX2的矿用巡检机器人整体设计方案,并重点分析了其中的行人检测模块的设计方法;然后提出了一种基于改进SSD(Single Shot Multi Box Detector)的井下行人检测算法,优化了SSD网络模型中各预测层中默认框(Default Box)的宽高比及个数,设计了一种轻量级特征提取网络以及一种基于残差的预测模块,自制了井下行人数据集,该算法在井下行人测试集中的平均准确率为88.9%,在嵌入式平台Jetson TX2中检测速度为14fps;为了进一步提高算法在Jetson TX2中的检测速度,利用Tensor RT推理加速技术对算法进行压缩优化,经过加速优化后算法的平均准确率为87.2%,检测速度为48fps,并在实验室环境下对加速后的算法进行了测试,结果表明该算法能够有效检测出视频区域内的行人目标,并给出相应的报警提示,实现了对视频内行人的实时检测,但对于距离较远以及尺度较小的行人会有轻微的漏检现象。本文基于改进SSD的井下行人检测算法可应用于矿用巡检机器人中实现对井下工作人员的自动检测,也可应用于井下定点监控以及井下机车的辅助驾驶系统,具备较高的应用价值。更多还原

【Abstract】 Mine inspection robot is an important auxiliary equipment in the field of safe production of coal mines,which is of great significance for improving the efficiency of inspection and ensuring the safety of underground production.At present,mining inspection robots are limited to the collection of surrounding environmental parameters and equipment status information,and do not have the function of real-time inspection of underground workers.At the same time,due to the adverse conditions of the underground environment,insufficient lighting,mixed image background,and poor image acquisition quality,traditional pedestrian detection methods cannot be directly applied to mining inspection robots,and it is difficult to achieve real-time detection and alarm of the dangerous behaviors(such as near the head,tail,tensioning device and other easy-to-clamp parts of the belt conveyor,across the conveyor belt,etc)for workers along the inspection line,as well as staff on-the-spot inspection functions.The purpose of this thesis is to research and propose a deep learning-based mine pedestrian detection algorithm,which is deployed in the embedded platform Jetson TX2,to realize the automatic detection of underground workers of mining inspection robots.Firstly,an overall design scheme of mining inspection robot based on Jetson TX2 was proposed,and the design method of pedestrian detection module was analyzed in detail.Then,based on the improved SSD(Single Shot Multi Box Detector)algorithm,a mine pedestrian detection algorithm was proposed to optimize the aspect ratio and number of the default boxes in each prediction layer of the SSD network model.We designed a lightweight feature extraction network and a prediction module based on residuals.Experimental results show that the algorithm has an average accuracy rate of 88.9% in the mine pedestrian test set and a detection speed of 14 fps in the embedded platform Jetson TX2.In order to further improve the detection efficiency of the algorithm in the embedded platform Jetson TX2,the algorithm was compressed and optimized using Tensor RT inference acceleration technology.The experimental results show that the average accuracy of the algorithm after acceleration and optimization is 87.2%,the detection speed is 48 fps.Then the accelerated algorithm was tested in the laboratory environment.The results show that the algorithm can effectively detect pedestrian targets in the video area and give corresponding alarm prompts to achieve real-time detection of pedestrians in the video.But pedestrians with longer distances and smaller scales may slightly miss detection.The mine pedestrian detection algorithm based on improved SSD proposed in this thesis can be used in mine inspection robots to automatically detect underground workers,and it can also be used in underground fixed point monitoring and auxiliary driving systems for underground locomotives,which has high application value.更多还原