摘要
目前,卷积神经网络(Convolutional Neural Network,CNN)模型在各种视觉任务中取得了巨大的成功,但行人检测方面的关键尺度问题仍有待进一步研究。为达到在交通场景下准确识别和定位小目标行人的识别与定位,提出了基于多尺度感知的改进Fast-RCNN模型,对Caltech行人数据集中的小目标行人图像进行检测。通过利用训练后的尺度感知权重,将大尺度子网络和小尺度子网络合并到统一的结构中,并利用对象建议的高度为两个子网络指定不同的尺度感知权重,同时将原模型中的VGG-16特征提取网络替换深度残差网络(ResNet-50)以获取更多特征。最后,对比所提改进模型和基础的Fast-RCNN的模型,发现所提模型行人识别准确率为97.49%,比未改进前提高了4.36%;再和传统的机器学习方法对比(基于HOG特征的SVM识别方法和基于ICF特征的AdaBoost识别方法),发现所提模型效果仍为最好。结果表明,该方法对交通场景下小目标行人的识别效果较好,能够为智能车辆图像识别系统和智慧交通提供参考。
At present,Convolutional Neural Network(CNN)model has achieved great success in various visual tasks,but the key scale of pedestrian detection still needs to be further studied.In order to accurately identify and locate small target pedestrians in traffic scenes,an improved fast RCNN model based on multi-scale perception is proposed to detect small target pedestrian images in Caltech pedestrian data set.By using the trained scale perception weight,the large-scale sub network and small-scale sub network are combined into a unified structure,and different scale perception weights are specified for the two sub networks by using the height suggested by the object.At the same time,the vgg-16 feature extraction network in the original model is replaced by the depth residual network(resnet-50)to obtain more features.Finally,comparing the improved model and the basic fast RCNN model,it is found that the pedestrian recognition accuracy of this model is 97.49%,which is 4.36%higher than that before the improvement;Compared with the traditional machine learning methods(SVM recognition method based on hog feature and AdaBoost recognition method based on ICF feature),it is found that the effect of this model is still the best.The results show that this method has a good effect on the recognition of small target pedestrians in traffic scenes,and can provide a reference for intelligent vehicle image recognition system and intelligent transportation.
作者
张鹏飞
蓝维旱
高峰
王迎旭
ZHANG Pengfei;LAN Weihan;GAO Feng;WAMG Yingxu(China Information Consulting&Designing Institute Co.,Ltd.,Nanjing 210019,China;Guangdong Communication Polytechnic,Guangzhou 510630,China)
出处
《通信电源技术》
2022年第1期87-90,112,共5页
Telecom Power Technology
