Fine-grained Vehicle Recognition In Traffic Surveillance

With the development of society and economic,the number of vehicles keeps in-creasing,which has brought much pressure to the management of transportation and public security systems.In order to maximize the use of existing traffic construction and provide reliable support for travel safety and transportation efficiency,it is im-perative to build a large-scale,comprehensive Intelligent Transportation System(ITS).Fine-grained vehicle recognition in traffic surveillance can identify the cate-gory,manufacture and model of the target vehicle,which can be applied in various scenarios,such as traffic flow monitoring,highway automatic charging,criminal tracking and investigation etc.,which makes the study of this paper significant to the establishment of modern ITS.In the task of fine-grained recognition,the major challenge lies in that the dif-ference between subcategories can be very subtle while the general appearances are highly similar.Traditional methods often extract basic features from the geometric sketch,license plate or frontal face of vehicles,which don't perform very well in the field of fine-grained recognition.In this paper,we propose a part-based method combining high-level and low-level feature for vehicle fine-grained recognition in traffic surveillance.The main contribution of this paper is as follows:1.For there are no available dataset for fine-grained vehicle models in traffic sur-veillance,we collect 4584 high-resolution vehicle images belongs to 50 common vehicle models of 8 makers.There is 5-9 types belong to each maker.There are at least 40 images of each type.All images are labeled with the form of "manufacture-model-year",and vehicle parts are manually annotated.2.We adopt a part-based method,in which seven semantic parts of vehicle appear-ance are defined empirically.And we modified strongly supervised deformable part model(SSDPM)to localize these parts.The high-level feature of parts is extracted by fine-tuning convolutional neural network(CNN),combing with the low-level feature based on the histogram of gradient(HOG),comprising comprehensive fi-ne-grained feature of the target vehicle.The experiments show the accuracy rate of vehicle recognition based on our proposed fine-grained feature is higher than tradi-tional methods based on low-level feature by 5.65%;3.As different parts are not equally important in recognition task,an entropy-based score is defined based on frequency of the part occurs in each sub-category to evalu-ate the discriminative ability of parts.After the initial recognition using single fea-ture of single part separately,a novel voting mechanism is performed by using nor-malized discriminative scores as weights to optimize the final recognition results.The experiments on our dataset show the accuracy rate of our method is higher than the state-of-art method by 3.1%-7.3%with lower complexity and higher efficiency.