OERFF: A Vehicle Re-Identification Method Based on Orientation Estimation and Regional Feature Fusion

Abstract

Vehicle re-identification (re-id) is an important issue in the transportation and vehicle tracking area. The existing FastReID framework has made significant improvements in data preprocessing, model structure and parameter configuration of re-id. However, the FastReID did not consider the directional differences between vehicle images, which makes it difficult to balance the image pairs with different directional differences during calculating the similarity between vehicles. To make a further improvement based on the FastReID, this paper proposes an Orientation Estimation and Regional Feature Fusion method, named OERFF. In OERFF, the orientation estimation model is trained to judge the orientation and the main region of a vehicle in an image. Then the dedicated region feature model is utilized to extract the regional feature of a vehicle according to its proper orientation. Finally, a feature fusion strategy is applied with weighted distance according to the orientation difference of vehicles to further improve the identification accuracy. Extensive experiments are conducted based on three workbench datasets and results show that the proposed method can improve the mAP by 2.5% on the VeRi-766 dataset and 3.2% on the VERI-Wild dataset. The accuracy of rank-1 is improved by 2.4% on the VehicleID dataset.

Keywords

• Deep learning
• vehicle re-identification
• vehicle re-id
• orientation estimation and regional feature fusion