基于深度学习的交通标志测距方法研究  

Research on Traffic Sign Ranging Method Based on Deep Learning

作　　者：王艳丽[1] 王志林[1] 侯宪春[1] 聂妍[1] 储怡[1] WANG Yan-li;WANG Zhi-lin;HOU Xian-chun;NIE Yan;CHU Yi(Jiamusi University College of Science,Jiamusi 154007,China)

机构地区：[1]佳木斯大学理学院,黑龙江佳木斯154007

出　　处：《电脑知识与技术》2020年第15期1-3,7,共4页Computer Knowledge and Technology

基　　金：佳木斯大学科学技术重点项目(Lz2014-007)。

摘　　要：自动驾驶技术和高级驾驶辅助系统设计成为近几年的研究热点,对交通标志进行实时的识别和定位在其中扮演了至关重要的角色,但目前研究大都局限于对交通标志的检测和识别,没有涉及对交通标志距离的测量。针对以上问题,提出了一种基于深度学习的交通标志检测与测距方法,该方法是基于回归的深度神经网络检测和步进式三维重建测距网络相结合,通过单阶段目标检测神经网YOLOv3完成交通标志的识别,并输出交通标志的像平面坐标,进行步进式转换处理,预测交通标志的距离。实验结果表明:检测速度达到29f/s(每秒传输帧数),距离检测平均误差为3%左右,准确率达到91.58%以上,该方法在精度和速度上均能获得较高的检测性能,完全满足自动驾驶中交通标志检测的实时性需求。The design of autonomous driving technology and advanced driver assistance systems has become a research hotspot in re⁃cent years.Real-time recognition and location of traffic signs have played a vital role in it,but most of the current research is limit⁃ed to the detection and recognition of traffic signs.No measurement of the distance of traffic signs is involved.A method of traffic sign detection and ranging based on deep learning is proposed to solve the above mentioned problems.This method is a combina⁃tion of regression-based deep neural network detection and stepping three-dimensional reconstructed ranging network,Recogni⁃tion of traffic signs through single-stage target detection neural network YOLOv3,and the image plane coordinates of the traffic sign are output,and with stepping conversion processing is performed to predict the distance of the traffic sign.The experimental results show that the detection speed reaches 29f/s(transmission frames per second),the average distance detection error is about 3%,and the accuracy rate is above 91.58%.The method can obtain high detection performance in accuracy and speed to meets the realtime needs of traffic sign detection in autonomous driving.

关 键 词：交通标志检测 深度学习 测距 YOLOv3 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]