Vehicle target detection based on R-FCN

Abstract

The traditional vehicle target recognition algorithm requires that the vehicle be artificially designed in different environment, and the detection time is long and the recognition rate is low.This paper proposes a method of the vehicle target detection based on region-based fully convolutional network(R-FCN). This method is based on fully convolution network idea of deep learning, applying R-FCN framework and combining vehicle database in ImageNet. In this method, online hard example mining (OHEM) is used to optimize the network parameters. After repeated iterations of the network training, R-FCN model of the vehicle target detection is finally obtained, and then the vehicle target is detected by R-FCN. Compared with method of the traditional vehicle detection, method of the vehicle target detection based on deep learning avoids the feature selection problem of the traditional detection and has obvious advantages in reducing the detection time and improving the vehicle recognition rate.Through the detection and identification of four main types of vehicles (buses, coupes, vans, suvs) in urban roads, experimental results show that this method achieves an average recognition rate of 87.48% for vehicle target detection. And the deformation of the the vehicle in the sample is samller, the vehicle target detection effect is better.