基于多元信息融合的车路协同重识别算法  

Re-identification Algorithm for Cooperative Vehicle-infrastructure System Based on Multi-information Fusion

作　　者：钱敏 耿可可 殷国栋[1] 李尚杰 王子威 孙宇啸 QIAN Min;GENG Keke;YIN Guodong;LI Shangjie;WANG Ziwei;SUN Yuxiao(School of Mechanical Engineering,Southeast University,Nanjing 211189)

机构地区：[1]东南大学机械工程学院,南京211189

出　　处：《机械工程学报》2024年第16期231-240,共10页Journal of Mechanical Engineering

基　　金：国家自然青年基金(51905095);江苏省重点研发计划(BE2019004);国家杰出青年科学基金(52025121);国家自然科学基金(51975118)资助项目。

摘　　要：针对车路协同系统的感知信息冗余问题,提出一种基于多元信息融合的重识别算法。算法输入为两个目标的图像与轨迹,通过计算图像相似度与轨迹相似度,并设计考虑多种权重的融合策略,最终得到两个目标的相似度,并做出对应的重识别处理。对于图像相似度,设计一种轻量化的图像相似度孪生网络;对于轨迹相似度,设计一种轨迹相似度算法;对于融合策略,通过计算图像质量权重、轨迹质量权重、图像相似度权重、轨迹相似度权重等四种权重,使用融合函数实现对图像与轨迹信息的融合。搭建真实交通场景下的车路协同系统,并基于该系统进行多种工况的试验来验证算法的有效性。结果表明,相较于SVDNet、Cam-GAN、MultiScale等基于单一信息元的重识别算法,基于多元信息融合的车路协同重识别算法提高了30%以上的重识别正确率和降低了20%以上的重识别漏检率。A re-identification algorithm based on multi-information fusion is proposed to solve the problem of perception information redundancy in cooperative vehicle-infrastructure system(CVIS).The inputs of algorithm are the images and trajectories of the two observed targets.By calculating the image similarity and the trajectory similarity,and designing a fusion strategy considering multiple weights,the overall similarity of the two targets can be obtained.Then,a re-identification process is made according to overall similarity of the two targets.For the image similarity,a lightweight image similarity Siamese network is designed;for the trajectory similarity,a trajectory similarity algorithm is designed;for the fusion strategy,four kinds of weights,including image quality weight,trajectory quality weight,image similarity weight,and trajectory similarity weight,are calculated.A fusion function is proposed to realize the fusion of image and trajectory information.A platform of CVIS is built in real traffic scenario.Based on this platform,a variety of experiments under different working conditions are carried out to verify the effectiveness of the algorithm.The results show that compared to the re-identification algorithms based on single information source such as SVDNet,Cam-GAN,MultiScale,the re-identification algorithm based on multi-information fusion improves the re-identification accuracy rate by more than 30%and reduce missed detection rate by more than 20%.

关 键 词：车路协同 重识别 图像相似度 轨迹相似度 多元信息融合 

分 类 号：U461[机械工程—车辆工程]