摘要
为了从广域视角准确提取道路交通信息,提出了一种用于无人机视频车辆自动识别的改进Faster R-CNN模型.该模型以基于ZF网络的Faster R-CNN为原型,优化调整学习策略、训练图像尺寸、学习率等模型参数,调整RPN网络卷积核并引入SoftNMS算法,增加1~3个特征提取卷积层和激活层.基于无人机交通视频构建了训练图像集,对现有Faster R-CNN模型及改进模型进行训练和测试.结果显示,与采用Step学习策略的模型相比,采用学习策略Inv的模型车辆识别平均准确率提高了0.4%~9.4%.引入SoftNMS算法的模型比引入前的模型平均准确率提高了0.1%~7.9%.提出的改进模型平均准确率为94.6%,较基于ZF的Faster R-CNN模型、基于VGGM的Faster R-CNN模型和基于VGG16的Faster R-CNN模型分别提高了13.1%、13.1%和4.1%,且训练时间减少约3%,对多种场景的视频车辆检测具有较好的适用性.
To accurately extract road traffic information from regional perspective,a modified Faster R-CNN(region-based convolutional neural networks)model was proposed for automatic vehicle detection with UAV(unmanned aerial vehicle)videos.Taking Faster R-CNN based on ZF(Zeiler and Fergus model)network as a prototype,the model parameters such as the learning strategy,the training image dimensions and the learning rate were optimized and adjusted.The RPN(region proposal network)kernels were revised and the SoftNMS(soft non-maximum suppression)algorithm was adopted.One to three convolutional layers and corresponding activation layers for feature extraction were added.The training image sets were constructed based on UAV traffic videos for training and testing existing Faster R-CNN models and modified models.The results show that the average vehicle detection precision of the models adopting learning strategy Inv is 0.4%to 9.4% higher than that of the models utilizing learning strategy step.The average vehicle detection precision of the models adopting SoftNMS algorithm is 0.1%to 7.9%higher than that of the models without SoftNMS.The average vehicle detection precision of the modified model achieves 94.6%,which is 13.1%,13.1% and 4.1%higher than those of the ZF based Faster R-CNN,VGGM(visual geometry group model with medium architecture)based Faster R-CNN and VGG16(visual geometry group model with 16 convolutional layers)based Faster R-CNN,respectively.The training time decreases by about 3%,and it has good adaptability to vehicle detection from videos in several scenarios.
作者
彭博
蔡晓禹
唐聚
谢济铭
张媛媛
Peng Bo;Cai Xiaoyu;Tang Ju;Xie Jiming;Zhang Yuanyuan(Chongqing Key Laboratory of Traffic System and Safety in Mountain Cities,Chongqing Jiaotong University,Chongqing 400074,China;College of Traffic and Transportation,Chongqing Jiaotong University,Chongqing 400074,China)
出处
《东南大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2019年第6期1199-1204,共6页
Journal of Southeast University：Natural Science Edition
基金
国家自然科学基金资助项目(61703064)
重庆市科委基础研究与前沿探索资助项目(cstc2017jcyjAX0473)
重庆市技术创新与应用示范资助项目(cstc2018jscx-msybX0295)
合肥工业大学城市交通管理集成与优化技术公安部重点实验室开放基金资助项目(2017KFKT01)
重庆交通大学山地城市交通系统与安全重点实验室开放基金资助项目(2018TSSMC05)
关键词
智能交通
车辆检测
深度学习
无人机视频
FASTER
R-CNN
intelligent transportation
vehicle detection
deep learning
UAV(unmanned aerial vehicle)videos
Faster R-CNN(region-based convolutional neural networks)
