Vehicle re-identification for Visual Surveillance

Abstract

Intelligent video surveillance systems are becoming an increasingly important major area of artificial intelligence as the use of motor vehicles in today’s transportation networks grows. In recent times re-identifying vehicles over a surveillance camera network have been proven to be the most effective method of efficiently controlling traffic, upholding the law, gathering information, and programming the traffic. Vehicle Re-ID seeks to recognize a target vehicle in several, non-overlapping camera perspectives. Compared to the com mon person re-ID issue, it has gotten far less attention in the com puter vision community. The absence of pertinent research data and the unique 3D structure of a vehicle are two potential causes of this delayed progression. The major concerns are the modest inter-class disparity caused by almost similar identities and the large intra-class distance based on a variety of circumstances, such as background interference, illumination, and viewpoints. In the past, the majority of the works focused on particular viewpoints (e.g. front and rear only) however, actual situations demonstrate that these methods are ineffective, as vehicles typically appeared from various perspectives to cameras. To this end, we present PAK Vehicle Re-ID Dataset (PV-ReID) with arbitrary viewpoints including unique vehicle iden tities based on Asian regions like Pakistan and India. In addition, we present a generalized pipeline for vehicle re-identification systems, which achieved 84.6%, 93.4%, and 78.3% accuracy, respectively. Lastly, a comprehensive comparison has been carried out to demonstrate the effectiveness of our proposed dataset and vehicle re-identification pipeline.