Vehicle re-identification is identifying and position referencing a vehicle of interest through images or video streams collected using a network of camera stations. This task is crucial in intelligent transportation systems and smart city development. However, vehicle re-identification poses a difficult problem due to various challenges, including intra-class variations, inter-class similarities, and near duplicate pairs. These challenges lead to a considerable false-positive rate in retrieved candidates. This research aims to design and develop a robust deep-learning-based vehicle re-identification system to address these challenges.

Effective training of deep learning models requires image down-scaling in preprocessing step, which can lead to a loss of detail in the input image. A recognition-aware attention-embedded learnable resizing module is employed in the preprocessing step to tackle this challenge. Furthermore, to overcome the inter-class similarity problem, a harder-negative-aware triplet loss, which produces a discriminative embedding space by focusing on dissimilarities in the dataset, is proposed. To improve the detailed-level perception of the framework, we incorporate the part-based local features into the feature extraction step along with the global information. The focus on the local features is extended further into the post-processing step, where the proposed candidates are validated via text-number tags (placed on the vehicle's body) verification for added robustness. Finally, to explore the effectiveness of the proposed vehicle re-identification framework, extensive experiments are conducted on a novel vehicle dataset. The dataset includes images of commercial vehicles and large trucks. The results show that our framework yields a Cumulative Match Curve at rank 1 of 99.62% and a Mean Reciprocal Rank of 99.56% on the novel dataset of commercial vehicles. Further analysis and study of the relationship between components of this framework point out the effect of the harder-negative-aware triplet loss on decreasing the false positive rate (4% improvement of Cumulative Match Curve at rank 1) by targeting the inter-class similarities challenge. Furthermore, the results display the exemplary impact of the local features extraction module on handling the near-duplicate challenge (14% further improvement of Cumulative Match Curve at rank 1) in the vehicle re-identification task.