Vehicle Re-identification in BEV Space with Deep Cosine Metric Learning

Re-identification (ReID) in autonomous vehicles faces limitations in image-based algorithms, highlighting the importance of sensor fusion in Bird’s Eye View (BEV) space to improve identification under adverse conditions. This study aims to develop a ReID algorithm for images that incorporates camera-to-BEV transformation, performing the re-identification process directly in the BEV space while retaining access to the visual features of the images in this representation. The algorithm development was divided into three stages: creating a camera-to-BEV transformation algorithm based on Lift-Splat-Shoot that preserves features in BEV; extracting a ReID dataset in BEV from Nuscenes using the transformation algorithm; and developing a ReID algorithm based on Deep Cosine Metric Learning, trained with the generated dataset. The results suggest that this approach can outperform image-based models in mAP for pedestrians and achieve similar performance for vehicles in mAP and also in Rank-1 and Rank-5 precision.