Faster-RCNN的车型识别分析

车型识别是目标检测领域在智能交通的重要应用,也是近年来国内外学者的研究热点之一。针对已有车辆检测方法缺乏识别车型能力的问题,提出了基于Faster-RCNN目标检测模型与ZF、VGG-16以及ResNet-101 3种卷积神经网络分别结合的策略,实验对比了该策略中的3种结合模型方案在BIT-Vehicle和CompCars2种大型车型数据库的车型识别能力。在BIT-Vehicle数据集上,基于Faster-RCNN与ResNet-101结合模型方案的车型识别率高与其余2种结合模型方案,其车型识别率高达91.3%;在迁移测试CompCars数据集上,3种结合模型方案均展现了很好的泛化能力。

Vehicle detection based on faster-RCNN

As one of the object detection, vehicle detection, which has been a hot research area in recent years, is one of the important application in intelligent transportation system. To figure out the problem that vehicle detection is lack of the ability of vehicle category recognition, we adopted the strategy of integrating the Faster-RCNN(region-based convolutional neural networks)model with 3 different convolutional neural networks (ZF, VGG-16 and ResNet-101)respectively. By comparing the vehicle category recognition results of the 3 integrating strategies on BIT-Vehicle database and CompCars database, the strategy integrating the Faster-RCNN model with ResNet-101 shows the best result among the 3 models and recognition accuracy reaches 91.3% on BIT-Vehicle database. On the migration test CompCars database, 3 strategy models show good generalization ability.