关键词: 车辆检测；卷积神经网络；识别率；锚框

Abstract: In recent years, solving traffic congestion has become the first task of traffic problems, and the application of vehicle detection and identification in many aspects has become one of the more commonly used solutions. Faster R-CNN based on convolutional neural network has gradually become an important method for image object detection and recognition. The algorithm has high detection accuracy and small limitations, so it has received widespread attention. At present, the commonly used algorithm models are based on the correction of the parameters of the model itself, and rarely involve improvements other than the network structure. In this paper, the network structure improvement, anchor frame extraction improvement and candidate frame parameter correction method based on faster R-CNN vehicle object detection method are discussed, especially the retraining after preprocessing of data sets is often more effective. These methods can not only make the network structure lightweight, but also increase the generalization ability of the network, so that Faster R-CNN can complete more complex traffic environment detection tasks, and the recognition accuracy has also been significantly improved.

Key words: Vehicle detection; Convolution neural network; Recognition rate; Anchor frame