在车辆重识别任务中,车辆视角的多变性会影响算法的准确性。 为了解决视角多变对重识别准确性的影响,本文提出 了一种基于局部特征与视点感知的车辆重识别方法。 首先,使用语义分割算法将车辆解构为正面、背面、侧面、顶部 4 个部分, 以提高车辆的细粒度表征。 通过设计一种车辆视点感知网络,来输出视点的预测概率信息,据此概率信息动态平滑地呈现车辆 视点感知效果。 利用视点感知效果,为车辆每个局部区域赋予不同的权重,达到缩短类内距离,扩大类间差距,减少视角变化对 车辆重识别的影响。 利用公开数据集进行实验,其中 VeRi776 数据集的 mAP 可达到 80. 9% 。 结果表明,本方法可有效提高车 辆重识别精度。 结合消融实验证明了视点感知的平滑表示对多视角下车辆重识别的有效性。
The change of vehicle view may affect the accuracy of the re-identification algorithm. To solve the influence of changing viewpoints on the accuracy of re-identification, we propose a vehicle re-identification method based on local features and viewpoint perception. First, a parsing module is trained to parse a vehicle into four different views, front, back, side, and top. In this way, the fine-grained representation of the vehicle is improved. Then, we intrduce a vehicle viewpoint-aware network. The output of the network is the predicted probability information of the viewpoint, and the vehicle viewpoint perception effect is dynamically and smoothly represented according to the probability information. Finally, the viewpoint-aware effect is used to assign different weights to each local area of the vehicle to shorten the intra-class distance, expand the inter-class distance, and reduce the impact of viewpoint changes on vehicle re-identification. This method is evaluated on public datasets, including VeRi776 and VehicleID. The accuracy of mAP on VeRi776 dataset has achieved 80. 9% . Experimental results show that the proposed method can effectively improve the accuracy of vehicle re-identification. Ablation experiments demonstrate the effectiveness of the viewpoint-aware smooth representation for vehicle reidentification from multiple viewpoints.