Vehicle Re-identification Of Based On Pose Guidance Method

With the construction of smart cities,intelligent traffic video surveillance has become an increasingly important core link.Vehicle Re-identification(Vehicle Re-ID for short)is an emerging technology in intelligent traffic video analysis and it is becoming a research hotspot.The main task of vehicle re-identification is to quickly retrieve and locate the target vehicle in the monitoring network of multiple non-overlapping coverage cameras.In real and open traffic monitoring scenes,uncontrollable factors such as changes in ambient light,low camera resolution,and occlusion make it difficult to accurately obtain vehicle license plate information from video.Vehicle re-identification without using license plate information faces huge challenges.For example,different vehicles may have similar appearances,and the same vehicle under different poses or viewpoints may have very different appearances.Focusing on these issues,this paper is inspired by related research work,and proposes a research method of vehicle re-identification based on pose guidance.The specific content of this article is as follows:1.We research on the method of extracting visual features that are robust to vehicle pose changes.We design a pose-guided visual model KPGEV and design a pose classifier based on key points of the vehicle skeleton to obtain vehicle pose's features and estimate pose's categories.Through observation,it is found that the visual feature distance between the same vehicles in different poses may be greater than the visual feature distance between different vehicles in the same pose,resulting in errors in identity discrimination.Therefore,we adopt a loss function PG-Triplet to mine the hardest vehicle's negative sample pairs in same pose and the hardest vehicle's positive sample pairs in different poses.It can guide the adaptive fusion of vehicle global features and pose's features and extract visual features that are robust to vehicle posture changes.This visual method on public datasets can reach more than60% on m AP index.2.We propose the KPGST method based on vehicle pose guidance,and optimize the spatio-temporal constraint model by mining the relationship between vehicle poses and spatio-temporal distribution.The design motivation of this method is based on the fact that the direction of movement of the vehicle when passing multiple cameras continuously should be consistent.The KPGST method combines the vehicle's pose and camera shooting direction to estimate the direction of vehicle movement,and guides the spatiotemporal constraint model based on the relative vehicle movement direction and the camera topology.On the public dataset,the effectiveness of the method is proved by comparative experiments.Moreover,we fuse the visual method and the spatiotemporal method based on Bayesian probability,and propose a unified re-identification framework.Through multiple comparisons with the results of existing research methods,this re-identification framework is verified to be efficient and advanced,and on the public dataset can reach more than75% on m AP index.Our research results have a positive effect on the field of intelligent monitoring and intelligent transportation.