Research And Application Of Vehicle Retrieval Based On Image Content

Intelligent detection and recording system for road vehicles is a kind of traffic monitoring system,which is set up in fixed areas to shoot and record the moving vehicles in real time.Currently,the traffic criminal and security cases are more and more frequent,especially escaping after the accident or robbery of motor vehicles.Traffic cases can be reduced by improving the intelligent performance of traffic monitoring system.And detect the suspected vehicles is the main research field of traffic monitoring system.Image retrieval technology is a new way to solve traffic problem,then establish a vehicle retrieval system to effectively find the suspect vehicle has a very practical value.The vehicle retrieval system can be determined by the vehicle's significant area feature information,such as ornaments,posters and other significant signs.Input the extraction of significant areas of the vehicle,then the algorithm can automatically identify the suspect that contain these objects.In this paper,there are two modules of the algorithm,the former one is vehicle detection based on convolution neural network frame;the latter one is image recognition based on BOW.And the research work and innovation of this paper are as follows:(1)Analyze and compare several common convolution neural networks frame,and do fine-turning of YOLO detection network.Using the road bayonet image set on the ground provided by traffic department,do train and test vehicle detection network model.In order to speed up the vehicle detection speed,but ensure the detection rate is not affected,making the following improvements: Use shallow network for transfer learning,use inception-v3 filtering overlay mechanism to increase fine grain size to improve vehicle detection and location performance.(2)Research the retrieval algorithm frame based on content-based image retrieval,especially analyze image feature extraction,feature quantification and the similarity measure.And analyze and compare several matching recognition algorithm.Finally the experiments show that the use of approximate nearest neighbor method based on visual dictionary to establish inverted index and binary feature fusion method can achieve better performance.And this paper proposes a kind of composite index(d-MI,double multiple index)framework to perform the fusion of SIFT feature and color feature based on the index level,and uses the embedded double binarization feature coding to improve the retrieval recognition accuracy,as the SIFT feature,only the local gradient distribution information is described.It is proved by experiment that the method not only speeds up the retrieval speed but also improve the retrieval accuracy.(3)In Addition,a feature combination database has been built for the suspect vehicle identification,and design and implementation a suspect vehicle retrieval system software based on d-MI.After the functional test and performance test,it shows that the improved algorithm can meet the actual requirements of vehicle detection and suspect target recognition.