面向公共安全监控的多摄像机车辆重识别

由于现有的车辆重识别方法大多是在已标注车辆边界框的图像间进行的,但在真实场景中无标注信息,同时环境的复杂性、车辆外观的相似性和多样性也是导致重识别精度不高的原因。因此,针对公共安全监控领域中无标注的原始视频,提出一种结合车辆检测与识别的多摄像机车辆重识别方法。首先设计了二值-单点多盒车辆检测网络以获取视频中的车辆边界框,并在线生成候选车辆数据库;其次设计了一种多任务孪生车辆识别网络以提高重识别精度;最后组建“VeRi-1501”车辆数据集。该数据集在现有数据集上扩充车辆身份,并均衡每个车辆身份在不同摄像机下的图像数量。该方法在VeRi-1501数据集和实际交通场景中识别准确且精度高。

Vehicle re-identification by multi-cameras for public security surveillance

The existing vehicle re-identification (Re-ID) methods mostly perform Re-ID between images marked with vehicle bounding boxes, but there are no vehicle bounding boxes in the real scene; at the same time, the complexity of environment and the similarity and diversity in appearance among vehicles can also cause a low accuracy of Re-ID. Therefore, this paper proposes a multi-camera vehicle Re-ID method combining vehicle detection and recognition for the unmarked original video in the field of public safety surveillance. First, the Binary-Single Shot MultiBox vehicle detection network is proposed to obtain the vehicle bounding boxes and generate candidate database online. Second, a multi-task Siamese vehicle recognition network is designed to improve the Re-ID accuracy. Finally, the “VeRi-1501” vehicle dataset is established, which expands the number of vehicle IDs and balances the number of images for each ID in the case of different cameras. The proposed method has achieved good results in the VeRi-1501 dataset and the actual traffic scene.