Research On Fine-grained Vehicle Classification Based On Deep Multi-task Learning

With the rapid improvement of national economic level and the rapid development of road construction,the number of vehicles in Chinese cities has been increasing in the past decade.Traffic problem has become an important problem of urban management,which has brought serious influence to the development of urban society and economy.Intelligent transportation system is the important direction of future transportation system.Fine-grained vehicle classification is the key technology of intelligent transportation video analysis.The purpose of fine-grained vehicle classification is to identify the specific vehicle type given a frame of vehicle video image.There have been a lot of research achievements about vehicle fine-grained classification,but it is still a difficult problem to be solved,facing the challenges of high intra-class differences and low inter-class differences.In order to improve the accuracy of fine-grained vehicle classification,this paper mainly studies fine-grained vehicle classification technology from the following three aspects:(1)In order to design a reasonable and feasible research route,this paper conducts a relatively extensive and in-depth literature survey on fine-grained vehicle classification,summarizes the main algorithms,and compares the performance of various algorithms on several influential fine-grained vehicle classification datasets.(2)In order to make full use of the viewpoint information of vehicles,a multi-task learning algorithm is constructed in this paper.The viewpoint estimation task is used to improve the precision of vehicle classification task.Experimental results on Boxcars and Compcars show that the proposed multi-task model has a greater improvement in accuracy than the baseline algorithm,and can better improve the accuracy of fine-grained vehicle classification by using geometric viewpoint estimation task.(3)In order to extract highly discriminant features,a feature extraction algorithm combining deep features and hand-crafted features is proposed in this paper.Specifically,in the hand-crafted featureconstruction,this paper firstly extracts the SIFT feature vector,and uses Fisher Vector algorithm to code the SIFT feature vector to get the FV-SIFT feature.Deep features are extracted using the Resnet network.Then bag-of-words encoding deep feature and FV-SIFT feature are used to get the fusion feature.Finally,SVM is used to classify and fuse the features.Experimental results on Stanford Cars-197 show that the proposed algorithm achieves better performance.