A Vehicle Re-ID Algorithm Based on Channel Correlation Self-attention and Lstm Local Information Loss

Abstract

Recently, the rapid development of vehicle re-identification (ReID) technology has facilitated the construction of intelligent transport systems. Mainstream ReID methods rely on the fusion of global and local features. In the global feature extraction, the channel attention modules are usually exploited in the network, most of which only focus on the channels’ importance and ignore the interactions among channels. In the local feature extraction, the additional annotation-based local feature extraction methods can focus on local information and improve the model’s performance but increase the workload of the data annotation and reduce the generalizability of the model. In this article, we put forward a new ReID Algorithm called CCSAM-LL. Firstly, a plug-and-play module based on channel correlation self-attention called CCSAM is introduced, which focuses on channel relevance and improves the characterization of global features. Secondly, we propose an Lstm-based loss, named LstmLocal loss, which takes into account local features without additional annotation. LstmLocal loss is trained with Triplet Hard loss and ID loss to improve the model’s ability to capture detailed features and accuracy in the retrieval task. Experimental results demonstrate that our approach outperforms the state-of-the-art methods on the challenging dataset VeRi776. Specifically, our approach achieves 83.18% mAP, 98.79% Rank5, and 48.83% mINP. The model is available at https://gitee.com/qitiantian128/ccsam-ll.

This work was supported by National Key Research and Development Program of China 2018YFB2101300 and the Dean’s Fund of Engineering Research Center of Software/Hardware Co-design Technology and Application, Ministry of Education, East China Normal University.