Recent Advances in Vision-based Vehicle Re-identification Datasets and Methods

Intelligent Traffic Systems (ITS) are currently one of the main concerns of governments and academia. Among the highly valued tasks of ITS is vehicle re-identification. Being able to detect unique IDs of vehicles in urban cities enables authorities to achieve active traffic control and forensic analysis of criminal activities. The abundance of low-cost surveillance cameras in residential areas makes it economically and feasibly attractive to study vision-based vehicle re-identification methods that doesn't depend on License Plate Recognition rather utilizes the visual appearance of vehicle to re-identify it in different non-overlapping cameras. The recent development of deep learning application in computer vision tasks encouraged researchers in the recent few years to utilize deep-learning vision-based methods to solve the vehicle re-identification problem. Also, that accelerated emerging of various vehicle re-identification datasets with different annotated attributes. We present six different datasets with different attributes of vehicles and conditions. We compare them quantitatively and qualitatively. We discuss to what extent each of them can resemble real-world scenarios. In addition, the vision-based vehicle re-identification task was approached using a wide range of approaches which made it complex for future researchers to choose a specific approach. Therefore, reviewing and making a semantic map of current methods, categorizing, and comparing them can be greatly helpful to future researchers. We categorized current methods into three main categories according to how it does the re-identification task: region-based, feature mapping, and similarity learning, and compare them regarding advantages, disadvantages concerning generalization, training stability and sensitivity, and target feature detection. We also discuss different methods in each category to review how development in each category was achieved through years.