Design And Implementation Of Video Vehicle Retrieval System Based On Vehicle Attribute Information

At present,most intelligent transportation systems supervise the behavior of vehicles in videos through license plates.However,the license plate may be blocked in a complex road background,poor camera quality and light may cause the license plate to be unclear,and even in the case where the license plate is unlicensed,license plate recognition cannot be performed.At this point,it is an important research direction to search the vehicle by its appearance.Distinguishing between two vehicles of the same color and model depends on the details of the body.However,for vehicles causing traffic accidents and suspected crimes,the traffic department or public security agency often have only relatively obscure pictures,or some witnesses ’semantic descriptions of the appearance of the vehicle.It is difficult to find the target vehicle directly by machine.In this case,the vehicle brands and models can be used to search for similar vehicle video surveillance,reduce the search range,providing directional information for the artificial investigation.Therefore,this thesis designs and implements a video vehicle retrieval system based on vehicle attribute information.On the basis of vehicle detection,the target vehicle in the video is retrieved through vehicle attribute information such as license plate or vehicle brand model.The main work includes:1)Aiming at the problem of inaccurate segmentation of license plate characters caused by license plate deformation and character adhesion,which leads to incorrect character recognition,a multi-output neural network is used to recognize seven characters at the same time for the entire license plate area.Train on the license plate data set constructed in this thesis,and test on the SYSU license plate data set.The results show that the accuracy of the license plate character recognition is 96.7%,and the detection time can meet the real-time requirements.2)Aiming at the problem that the intra-class differences of multi-angle fine-grained vehicle classification exceed the inter-class differences and it is not easy to distinguish,the network model is improved based on the study of fine-grained image classification,and Inception V3 B-CNN based on multi-branch outer product is proposed.Firstly,the methods based on fine-tuning and direct training is implemented on the three CNN models.The experimental results show that the Inception V3 method based on fine-tuning works best.Then,the Inception V3 based on fine tuning is used as the feature extractor of B-CNN,and use feature outer products for the six branches of its last inception module before performing feature stitching,which can also obtain more fine-grained while reducing network parameters combination characteristics of granularity.Finally,the method in this thesis achieves 96.72%accuracy on the Comp Cars dataset of 431 categories,and the detection speed is 45 ms per frame.3)A video vehicle retrieval system based on vehicle attribute information is designed and implemented.In view of the problem that the complex video background will interfere with license plate and vehicle model recognition,the Yolo V3 algorithm is applied to vehicle detection,and the image of the detected vehicle area is input into the above-mentioned license plate recognition or fine-grained vehicle classification model to obtain license plate or vehicle model semantics.After the feature is matched with the search condition,the target vehicle is retrieved.It is tested and evaluated in actual scenarios,and the results show that the system meets the expected requirements.