Vehicle Positioning Based On Vehicle Re-Identification And Pedestrian Navigation For Reverse Vehicle Searching Systems

The complex structure of large indoor parking lots and the low differentiation of different parking spaces make it difficult to find the target vehicle quickly.The reverse vehicle searching systems based on QR codes can locate vehicles by posting QR codes near parking spaces.However,users need to scan codes on their own initiative,which is not convenient and there is a high possibility that users don’t mark the location of vehicles.After locating the vehicle,it is necessary to reach the parking space quickly through navigation.Therefore,it is of great practical value to realize the non-inductive automatic vehicle positioning method so that users can find the vehicle quickly by pedestrian navigation.According to the structural characteristics of the indoor parking lot,a reverse vehicle searching frame that utilizes the deployed surveillance camera is proposed,which realizes the non-inductive vehicle positioning function through the re-identification of the monitoring images,and realizes the fast reverse vehicle searching through the pedestrian navigation.Firstly,based on analyzing the characteristics of reverse vehicle searching application scenarios,data acquisition and data augmentation are carried out,and the optimization direction of the vehicle re-identification method is analyzed from two aspects of accuracy and speed.Aiming at the problem of intra-class difference caused by perspective change,a multi-granularity vehicle weight recognition algorithm was designed based on an analytic view-perception network.According to the characteristics of a relatively simple classification task phase ratio and shorter inference time,different vehicle images are prefiltered through a coarse-grained classification network to reduce the number of images that need to be processed by a fine-grained view perception network,to improve the efficiency of view perception network for gender reidentification.Based on the spatio-temporal information of vehicle monitoring images,reordering was carried out by k-inverted nearest neighbor and spatio-temporal constraints to improve the accuracy of vehicle re-recognition and obtain the accurate parking area.The received bluetooth signal is spatio-temporal weighted buffered to improve the positioning accuracy and stability of bluetooth,to achieve high precision pedestrian navigation under the pre-deployed bluetooth beacon.The cumulative error of inertial navigation can be reduced by map constraint,and pedestrian navigation can only rely on the built-in sensor of mobile phones.The experimental results of Ve Ri776,a public vehicle recognition dataset,show that the improved algorithm can reduce the processing time by 12.5%～59.9% and the accuracy loss is less than 1% when the category is less than 5.Based on the Django framework which integrates the functions of video streaming,license plate recognition,vehicle detection,and vehicle re-identification,the web application of vehicle positioning on the management end was realized.The road network planning and model rendering were carried out with the 3D map model to realize the We Chat applet for pedestrian navigation on the client end.Using the developed web application and We Chat applet,it can realize the non-inductive vehicle positioning and the fast reverse vehicle searching.