Research On Vehicle Re-identification Algorithm Based On Multi-task Learning

In recent years,with the rapid growth of surveillance cameras,the monitoring data on vehicles has exploded,and a technology is urgently needed to analyze and process massive video data.The vehicle re-identification task has played an important role in building smart transportation and building a safe city because of its high efficiency,convenience and innovation in the efficient processing of massive data.The feature extraction of the vehicle is crucial for solving the vehicle re-identification task.At present,the traditional manual feature extraction method is highly susceptible to environmental factors such as illumination changes and camera angle settings,which makes it difficult to extract discriminative feature.Furthermore,many methods focus on feature extraction,ignoring its ranking optimization of the target after feature extraction.In this paper,a multi-task vehicle re-identification algorithm framework based on fusion ranking optimization method is proposed,which includes a multi-task learning based convolutional neural network model and a re-query ranking optimization algorithm.The details are as follows:(1)Vehicle re-identification algorithm based on multi-task learning.Existing vehicle re-identification methods mainly focus on extracting target features through a single convolutional neural network model.The commonly used single network models are the verification model and the identification model.Due to the difference in the loss function,they have their own advantages and limitations(the annotated information of the image cannot be fully utilized).Therefore,our method combines the advantages of the above two mainstream models,and simultaneously uses the similarity information and category information existing between the images to extract more robust image features.The algorithm is based on a Siamese network model with two branches,each of which can be thought of as a separate identification model,and finally merges the two branches for category validation.In other words,the network model performs three tasks simultaneously,including two classification tasks and one verification task.Experiments show that the features extracted by the multi-task network model are more robust than the single network model,which makes the accuracy of vehicle re-identification higher.(2)Ranking optimization method based on re-query.After using the deep learning method to extract vehicle features,many algorithms only use simple distance metric learning to sort the re-identification results,but it cannot guarantee high re-identification accuracy.The ranking optimization method proposed in this paper optimizes the final re-identification result by querying gallery many times and establish a double similarity relationship of fusion rankings.The similarity relationship includes:1)The similarity between the images from probe and the images from gallery;2)The similarity between images in gallery with strong similarity to the probe.Furthermore,combining the above-mentioned similarity relationship and the initial ranking of the query image,which are used to optimize the re-identification result.The experiment proves that the method can improve the accuracy of vehicle re-identification.(3)Multi-task vehicle re-identification algorithm framework based on fusion ranking optimization method.In this paper,the above two optimization methods are combined and a multi-task vehicle re-identification algorithm framework is proposed.Experiments are carried out to demonstrate that the good performance of our algorithm framework with improvements of 1.5%and 20%at the top-1 ranking matching accuracy on two mainstream datasets VeRi and VehicleID.