Research On Vehicle Retrieval Technology In Monitoring Video Based On Multi-feature Fusion

With the exponential growth of China's car ownership,urban traffic management has brought tremendous pressure.How to reasonably and effectively reduce the occurrence of traffic accidents and improve the efficiency of vehicle accountability after accidents is an urgent problem to be solved by the relevant government management departments.In the aspect of road management,intelligent traffic video surveillance system is the core focus of the work of each city traffic management department.Using these high-quality video data and advanced modern image processing and retrieval technology,an efficient supervisory query retrieval system emerged as the times require.Based on this,this paper designs and implements a vehicle retrieval system based on multi-feature fusion using vehicle information in urban road surveillance video.In this paper,the features of vehicle images in surveillance video are analyzed,and the extraction methods of high-level semantic features of vehicle images are deeply studied.Vehicle detection and vehicle association functions of video input module are designed to realize the extraction of low-level semantic features and deep-level semantic features of feature extraction module.In this paper,the validity of vehicle information extraction is validated by designing effective comparative experiments.The research results include the following work:(1)The video input module of vehicle system is realized,and the structure and principle of convolution neural network are fully studied.In the video input module of the system,two tasks of vehicle detection and Vehicle Association Based on Faster-RCNN convolution neural network are accomplished.The original vehicle detection algorithm is improved to improve the detection accuracy.An improved Fast-match fast template matching algorithm is proposed to calculate the template matching of special symbols such as vehicle annual inspection signs in the process of vehicle image retrieval.The improved Fast-match fast template matching algorithm retains the advantages of high robustness and fast operation of the original algorithm,and adds color features to the matching process,which greatly increases the accuracy of template matching.At the level of vehicle user level feature utilization,the algorithm construction and experimental verification are realized.The comparison experiments on two large data sets of vehicles provide validation.(2)The coarse-grained and fine-grained feature extraction of vehicle image in feature extraction module is realized.K-means clustering method is used to extract vehicle type features and coarsegrained features of vehicle color to complete the rough classification of the system.Deep hash network is designed and constructed to extract deep semantic features of vehicle images,and its high-dimensional features are coded into compact binary codes.The similarity measure criteria of feature vectors are constructed.(3)The design completes the fusion calculation of multi-features.According to the different retrieval environment,the vehicle in the sample database can be queried by using the input image database or the input monitoring video.The Vehicle Association retrieval in the traffic closed-loop area is completed through the system's picture output.