Research On Suspected Vehicle Retrieval Based On Multi-Feature Fusion

With the increasing diversification and intellectualization of crime,using advanced science and technology is the inevitable trend of modern public security case handling,which can improve the case handling efficiency,save manpower and material resources.The retrieval of suspected vehicle is one of the important tasks of video surveillance or image surveillance.Its goal is to find the vehicles from the massive video or image data which with the same identity as the target suspect vehicle,and then analyze and master the important clues,such as the suspect's driving track,life rule,correlation region and so on.Due to the low efficiency and low reliability of traditional manual vehicle retrieval,it is urgent to realize automatic suspected vehicle retrieval system.This paper takes the content-based vehicle retrieval technology as the research objective,aiming to improve the retrieval effect by optimizing image feature.The main work of this paper includes:(1)A vehicle 2D structural modeling based on window location and viewpoint normalization method is proposed.To ensure the high consistency of features in the expression of images from different angles,the vehicle viewpoint normalization method unifies the vehicle layout.A two-stage window positioning method is implemented since the window position is an important information to obtain the 2D structural model of the vehicle.(2)New ideas for solving various difficulties in vehicle color recognition are proposed,including fusion of various image preprocessing algorithms for improving the image color deviation,vehicle 2D structural model for obtaining the vehicle's color recognition region with invariant perspective,a more expressive HSV color feature based on super pixel.(3)A multi-task vehicle re-identification network MT-IVSN is proposed.On the base of existing research on re-identification network,with the specific attributes,sample similarity relation and other information of vehicles,a multi-task vehicle weight recognition network is proposed that can extract multi-attribute related fusion features.