基于SSD算法的电力巡线机器人障碍物识别  

Obstacle identification of transmission line inspection robot based on SSD

作　　者：毛先胤 刘宇 马晓红 张迅 王云 鲁彩江 MAO Xianyin;LIU Yu;MA Xiaohong;ZHANG Xun;WANG Yun;LU Caijiang(Guizhou Electric Power Test Research Institute,Guiyang 550002,China;School of Mechanical Engineering,Southwest Jiaotong University,Chengdu 610031,China)

机构地区：[1]贵州电网有限责任公司电力科学研究院,贵阳550002 [2]西南交通大学机械工程学院机电测控系,成都610036

出　　处：《自动化与仪器仪表》2020年第5期45-48,共4页Automation & Instrumentation

基　　金：国家自然科学基金(No.61801402);中央高校基本科研专项资金(No.A0920502051822-2);中国南方电网有限责任公司科技项目(No.GZKJXM20171600)。

摘　　要：电力巡线机器人常常工作于野外,周围有树木等环境干扰。同时,野外环境下,早晚太阳的照射角度和光照强度都有很大变化。而机器人在运动过程中,障碍物距离机器人的距离也在发生变化,这样会造成障碍物在图像空间中的尺度发生变化。这些光照变化和障碍物尺度的变化都会对识别的准确率造成巨大的挑战。针对这些挑战,提出采用深度学习领域的Single Shot Multi-Box Detector(SSD)算法来识别电力线障碍物。SSD算法可以自学习出对光照不敏感的特征,同时不同尺度的feature map可以识别不同尺度下的障碍物,可以克服光照和尺度变化带来的影响。由于SSD网络深度较深,容易出现梯度消失问题,在原有的SSD算法的基础上,在卷积层后面增减了Batch Normalization层,从而避免了梯度消失,同时加快了网络收敛速度。Power line inspection robots often work in the wild,surrounded by trees and other environmental disturbances.At the same time,in the wild environment,the angle of illumination and the intensity of the sun in the morning and evening change greatly.When the robot is moving,the distance of the obstacle from the robot is also changing,which will cause the scale of the obstacle in the image space to change.Such changes in illumination and obstacle scales can pose significant challenges to the accuracy of recognition.In response to these challenges,this paper proposes the Single Shot Multi-Box Detector(SSD) algorithm in the deep learning domain to identify power line obstacles.The SSD algorithm can self-learn features that are insensitive to illumination,while feature maps of different scales can identify obstacles at different scales,which can overcome the effects of illumination and scale changes.Due to the deep depth of the SSD network,the gradient disappearing problem is easy to occur.Based on the original SSD algorithm,the Batch Normalization layer is added after the convolutional layer,which avoids the gradient disappearing and speeds up the network convergence.

关 键 词：机器视觉 巡线机器人 障碍物识别 SSD 

分 类 号：TP24[自动化与计算机技术—检测技术与自动化装置]