基于区域卷积网络的行驶车辆检测算法  

Research on Running Vehicle Detection Algorithm Based on Regional Convolution Network

作　　者：曹长玉 郑佳春[2] 黄一琦 CAO Changyu;ZHENG Jiachun;HUANG Yiqi(Navigation College,Jimei University,Xiamen 361021,China;Information Engineering College,Jimei University,Xiamen 361021,China)

机构地区：[1]集美大学航海学院,福建厦门361021 [2]集美大学信息工程学院,福建厦门361021

出　　处：《集美大学学报（自然科学版）》2019年第4期315-320,共6页Journal of Jimei University：Natural Science

基　　金：福建省科技计划重点项目(2017H0028);福建省自然科学基金项目(2013J01203、2015J01265)

摘　　要：为解决多种天气与多种场景下主干道路行驶车辆检测存在的实时性、泛化能力差、漏检、定位不准确等问题,研究了基于TensorFlow深度学习框架的区域卷积神经网络(Faster R-CNN)算法,通过引入VGG16神经网络模型,优化ROI Pooling Layer,并采用联合训练方法,得到改进的算法模型。采用UA_CAR数据集进行模型训练,实现行驶中的车辆检测,测试结果与优化前Faster R-CNN比较,MAP提高了7.3个百分点,准确率提高了7.4个百分点,检测用时0.085 s,提高了对多种环境与场景的适应性。In order to solve the issue of poor real-time performance,poverty generalization ability,missed detection and inaccurate location of running vehicle detection on currently main roads in various weather and various scenarios,the regional convolutional neural network(Faster R-CNN) algorithm based on TensorFlow framework of deep learning is studied.By introducing the VGG16 neural network model,optimizing the ROI Pooling Layer,and adopting method of joint training,an improved algorithm model is obtained.The UA_CAR database is used for model training to carry out vehicle detection in the course of driving.Compared with the Faster R-CNN before optimization,the test results show that the MAP is increased by 7.3 percentage points,the accuracy rate is increased by 7.4 percentage points,and the detection time is 0.085 s.The network improved adaptability to multiple environments and scenarios.

关 键 词：行驶车辆检测 卷积神经网络 联合训练 

分 类 号：U495[交通运输工程—交通运输规划与管理]