Vehicle Re-Identification Collaborating Visual and Temporal-Spatial Network

Wenhua Fang,Jun Chen,Chao Liang,Yimin Wang,Ruimin Hu
DOI: https://doi.org/10.1145/2499788.2499844
2013-01-01
Abstract:Vehicle re-identification, retrieving a vehicle detected by one camera with the same vehicle by another camera, is an important problem in the video investigation application which is a technology for criminal investigation. In this task, it not only needs to classify the vehicle category, but also to identify a specific object in the category. Previous methods mainly focus on the vehicle categorization, which cannot identify the specific vehicle. In this paper, a two-stage strategy is proposed to accomplish vehicle re-identification in realistic surveillance videos. Specifically, in the first stage, a part-based appearance model fusing multiple visual features is proposed to represent the vehicle object, and then a coarse ranking list is generated by comparing appearance models of the probe and gallery vehicles. In the second stage, the temporal-spatial relation is introduced to re-rank the above visual-based ranking list, where vehicles of the same category and reasonable spatial-temporal relations are placed in top positions while those of mismatched types or relations are placed in rear positions. Both quantitative and qualitative experiments conducted on a real world dataset have validated the effectiveness of the proposed method.
What problem does this paper attempt to address?