基于Resnet-101深度学习网络的路面类型识别研究

Research on Road Terrain Recognition based on Resnet-101 Deep Learning Network

近年来,智能驾驶汽车已成为智能工业时代较有前景的发展方向,而作为智能车的关键技术的重要组成部分及重要研究方向之一,环境感知能为智能汽车的加速度控制、轨迹控制、燃油经济性的改善、安全性和舒适性的提高等方面提供重要依据。到目前为止,相较于由人工算法来提取的HOG、LBP等特征,深度学习能够提取到较为丰富的非人工提取的特征。本文基于迁移学习,通过Resnet-101深度学习网络实现对路面类型进行识别,训练所得的模型对本文给出的四种路面类型的识别准确率可达98.33%,相较于使用此前较为常用的Alexnet、GoogLenet等网络有着较为明显的提升,在相同的迭代次数的条件下,Resnet-101的准确率较高,且在道路实车试验中取得了良好的表现。

In recent years,intelligent vehicle has become a more promising development direction in the era of intelligent industry.As an important component of the key technology of intelligent vehicle and one of the important research directions,environment perception can provide important basis for acceleration control,trajectory control,improvement of fuel economy,improvement of safety and comfort of intelligent vehicle.Up to now,compared with HOG,LBP and other features extracted by artifi cial algorithms,depth learning can extract more abundant non-artifi cial features.Based on transfer learning,this paper realizes the recognition of road surface types through Resnet-101 deep learning network.The recognition accuracy of the four road surface types given in this paper can reach 98.33%with the training model.Compared with Alexnet and GoogLenet,which are commonly used before,this model has a more obvious improvement.Under the condition of the same number of iterations,the accuracy of Resnet-101 is higher,and it has achieved good performance in the road test.