Vehicle re-identification by fusing multiple deep neural networks

Vehicle re-identification has become a fundamental task because of the growing explosion in the use of surveillance cameras in public security. The most widely used solution is based on license plate verification. But when facing the vehicle without a license, deck cars and other license plate information error or missing situation, vehicle searching is still a challenging problem. This paper proposed a vehicle re-identification method based on deep learning which exploit a two-branch Multi-DNN Fusion Siamese Neural Network (MFSNN) to fuses the classification outputs of color, model and pasted marks on the windshield and map them into a Euclidean space where distance can be directly used to measure the similarity of arbitrary two vehicles. In order to achieve this goal, we present a method of vehicle color identification based on Alex net, a method of vehicle model identification based on VGG net, a method of pasted marks detection and identification based on Faster R-CNN. We evaluate our MFSNN method on VehicleID dataset and in the experiment. Experiment results show that our method can achieve promising results.

I. Introduction

Vehicle has attracted massive focuses in computer vision research field. There is a growing requirement of vehicle reidentification (Re-ID) and retrieval from large scale surveillance image and video database in public security systems. Fig. 1 gives us a description of this task. License plate is widely used as a unique ID of a vehicle, and license plate recognition has already been used in transportation management applications. However, in some cases the information of license cannot be used. Therefore, vision-based vehicle re-identification has a great practical value in realworld surveillance applications. Specifically, vehicle reidentification aims at identifying the same vehicle through different camera views.