摘要
由于不同道路监控视角下的车辆姿态不断变化,因此车辆重识别仍是智慧交通系统中一项具有挑战性的任务。现有的车辆重识别的方法大多数基于车辆的外观属性,但识别受光照和角度等因素影响导致识别效果较差。因此,设计了一种车辆姿态感知注意力增强网络以提高车辆在光照和角度等因素影响下的重识别效果。首先,将图片输入到卷积姿态网络中生成12个关键点重建车辆姿态信息,然后将输入图像车辆与目标图像车辆进行比较,提取出两辆车公共区域的特征;最后,计算车辆全局特征和局部特征之间的距离,并根据最终结果对识别结果进行排序。本文在Vehicle ID和Ve Ri776数据集上进行验证,实验结果表明,所提出的网络相较于其他模型top10的检测准确率提高了10%左右。
Owing to the continuous changes in vehicles under different road monitoring perspectives,vehicle reidentification is still a challenging task in intelligent traffic system.Most of the existing vehicle re-identification methods are based on the appearance attributes of the vehicle,but the recognition is affected by factors such as illumination and angle,which leads to poor recognition results.Therefore,this paper designs a vehicle posture perception attention enhancement network to improve the re-identification effect of vehicles under the influence of factors such as illumination and angle.First,input the image to the convolutional pose machine to generate 12 keypoints to reconstruct the vehicle frame,and then compare the input image vehicle with the target image vehicle to extract the features of the intersecting area between two images.Finally,the global distance and local loss of vehicle features are calculated,and the recognition results are sorted according to the final results.This paper verifies on Vehicle ID and Ve Ri776 data sets.The experimental results prove that the top10 detection accuracy of the proposed network is increased by about 10%than other models.
作者
朱肖磊
吴训成
Zhu Xiaolei;Wu Xuncheng(School of Mechanical and Automotive Engineering,Shanghai University of Engineering Sciences,Shanghai 201620,China)
出处
《电子测量技术》
北大核心
2021年第24期91-97,共7页
Electronic Measurement Technology
关键词
关键点
车辆姿态
注意力机制
车辆重识别
keypoints
vehicle posture
attention mechanism
vehicle re-identification
