卷积神经网络在车辆识别系统中的应用

Application of Convolution Neural Network in Vehicle Identification System

针对工厂厂房和仓库无人门禁系统需对货运车辆单独检测与识别的特殊需求,提出一种融合Darknet19网络与SSD(Single Shot-multibox Detector)模型的车辆检测与识别模型。首先,采集真实场景中包括行人、叉车、货车的大量图片并进行人工标注,构建一个私有数据集;其次,在Caffe框架下使用ImageNet2012数据集重新训练Darknet19网络,并通过更换基础分类网络及在每个卷积层后加入批归一化(Batch Normalization)层等方式改进SSD目标检测模型,构建出一个新的端到端的车辆检测模型。结果表明,该模型对货运车辆的平均查准率可达99.2%,检测帧率可达72帧/s,准确率与实时性均满足厂区环境检测货运车辆的要求。

In the view of the special requirements of separate detection and identification for vehicles of the unmanned access control system of factory buildings and warehouses, a vehicle detection and identification model based on Darknetl9 network and SSD （single shot-multibox detector） model was proposed. Firstly, by collecting a large number of images in the real scene including pedestrians, forklifts, trucks and manual labeling, a private data set was constructed. Secondly, the dataset of ImageNet2012 was employed to retrain the Darknetl9 network under the framework of Caffe. Besides, the SSD target detection model was improved by replacing the basic classification network and adding a batch normalization layer after each convolution layer to construct a new end- to-end vehicle detection model. The results show that the detection accuracy of the improved algorithm for truck can reach 99.2%, the detection speed can reach 72 frames/s, and the accuracy and real-time performance can meet the requirements of vehicle detection in storage environment.