| With the widespread application of vehicle monitoring video in intelligent transportation systems,it is worth discussing in the field of security to realize accurate and efficient retrieval of target vehicles in mass traffic surveillance video.Due to factors such as the natural environment and surveillance equipment,blurred,low quality images can be captured,for which existing plate detection algorithms can make errors in extracting plate information.At the same time the proliferation of video data volumes also makes centralised cloud computing a serious waste of resources and unable to achieve efficient retrieval of the target video.To solve the above problems,this thesis takes advantage of the advantage that edge computing can process data near the edge and reduce the burden of cloud computing.A vehicle videoed storage and retrieval technology based on edge computing is proposed to improve the precision of license plate recognition while reducing the force on cloud computing,and realize the rapid retrieval of massive cloud-stored vehicle surveillance videos.The main work and achievements of this thesis are as follows:Firstly,in order to relieve the calculation pressure on the central cloud server and improve the license plate location and recognition capabilities of surveillance videos,the vehicle picture is processed at the edge end by the means of deep learning.In the betterment of the license plate location algorithm,the Mask R-CNN is used to initially process the video key frames to dispose of the problem that the small target in the picture cannot be positioned correctly.Canny-Color algorithm is used to detect the edge of the license plate image,and remove the part outside the edge of the license plate and correct the license plate.When improving the license plate recognition technology,the Hyper LPR recognition model is modified according to the structure of Inception v1.This method can solve the problem of character segmentation required for license plate recognition.Secondly,in order to enhance the cloud storage ability,the compressed video is usually stored in the cloud,which brings about low recognition accuracy of the compressed vehicle surveillance video and is not conducive to subsequent video recovery.The video metadata can be used to describe the video content so as to solve the above problem and the information will not be damaged and lost by video compression.Vehicle information identified on board is stored as metadata in the video,then integrated into the video and transmitted to cloud storage.Thirdly,in order to decrease the action delay of video surveillance system and upgrade the retrieval efficiency of vehicle video.The video metadata values containing vehicle information and video address stored in cloud storage video by key-value pairs are taken out,and the metadata values are reserved in the database according to the form of hierarchical metadata.Based on the basis of self-constructed vehicle search terms,a word segmentation algorithm combining Bi-MM(Bidirectional Maximum Matching)and2-Gram is used to segment the search information,and realize the fast retrieval of vehicle surveillance videoed by metadata retrieval method. |
