深度预警网络的车辆识别与检测

Vehicle recognition and detection based on deep early warning network

针对驾驶过程中目标车辆实时检测准确度问题,提出一种基于YOLO和Darknet-19等融合式算法设计的深度预警网络,旨在实现对目标车辆的特征提取。网络基于车载单目摄像机拍摄的实时场景照片进行测试数据集的制作,预训练采用PASCAL VOC 2007数据集。应用于夜间可见度低的情况下,网络测试结果较YOLO网络准确率提高了4.89%,召回率提高了18.02%;应用于白天可见度高的情况下,准确率提高了8.72%,召回率提高了7.79%;平均检测时间也有所下降,实现了目标车辆检测准确度及速率的提高。

Aiming at the problem of real-time detection precision of target vehicle in driving process,a deep early warning network was designed based on the YOLO and Darknet-19 fusion algorithm,which will be mainly for the feature extraction of target vehicle.The network was based on the real-time scene photos taken by the vehicle monocular camera to produce the test data set.The PASCAL VOC 2007 data set was used for pre-training.Compared with YOLO Network which is commonly used in target detection at present,the accuracy and recall rate of YOLO Network are improved by 4.89%and 18.02%respectively when used in low night visibility.When applied to the situation of high daytime visibility,the precision and recall rate are increased by 8.72%and 7.79%,respectively.The average detection time is also reduced,which achieves the improvement of the detection precision and speed of the target vehicle.