Research On Key Technologies Of Vehicle Detection And Its Fine-grained Classification Based On Deep Learning

With the improvement of the status of intelligent transportation in the smart city,intelligent video surveillance system has received greater attention.How to make the video surveillance system have more intelligent functions has become one of the issues that people in industry and academic consider more and more.At present,most of the intelligent video surveillance systems only have functions such as the detection and tracking of objects,abnormal alarm and so on.While the core function of vehicle detection and fine-grained classification for intelligent traffic video monitoring systems has not been available as a product due to much more difficulties,especially the fine-grained classification function of vehicles,there has been no research results that can be put into practical.Aiming at the traffic scene image with complex background,this dissertation proposes a method based on deep learning to solve the problem of vehicle detection and fine-grained classification.Three key technical problems are proposed and partially solved,such as how to construct a large-scale vehicle dataset for training of deep learning models,how to construct a suitable deep learning model that can quickly and effectively detect the whole vehicles in images with complex background,and how to handle the issue that inner-class difference usually exceeds intra-class difference which make fine-grained classification of vehicle hard to tackle.Firstly,an integrated framework for vehicle detection and its fine-grained classification based on deep learning is proposed.When a traffic image with complex background inputted,this framework can first detect all the vehicles in this image.All detected vehicles with only a little background will improve the speed and accuracy of fine-grained classification to a great extent.Secondly,through the use of collaborative tagging network,making photos of vehicle model and automatic generation of vehicle images by three-dimensional model,a large-scale vehicle datasets can be constructed,which can save a large amount of image collection and annotation time,avoid marking inaccurate effect caused by this method.The experimental results show that the vehicle dataset collected by the above methods can be used as the effective data for the training of deep learning network modelThirdly,on the basis of Faster R-CNN generic object detection framework,according to the characteristics of rigid and symmetry of vehicles,this dissertation focuses on the improvement of RPN network in Faster R-CNN,proposes a vehicle detection network model which is more suitable for vehicles.The model comes up with the correct rate of 85%and 5 images per second of the detection speed.Finally,this dissertation put the car into 13 different parts,and for each part a part detector has been trained.Then a fusion of 13 part detectors forms the fine-grained classification network which can judge the classification of the input vehicle image.In this way,the problem of vehicle fine-grained classification changes to an easier way that we use parts of vehicle to distinguish different vehicle types.Experiments show that,68%accuracy can be achieved in five types of vehicles.