Key Technology Of Vehicle Re-identification In Complex Enviroments

Traffic video surveillance systems for complex scenarios have been acting an important role in road safety enforcement and intelligent traffic management.Vehicles are important behavioral subjects and main monitoring targets in traffic systems,and vehicle related research,such as vehicle tracking,vehicle detection,license plate recognition,vehicle re-recognition,etc.,has always been a research hotspot and focus in computer vision.Vehicle re-identification aims at quickly identifying and matching vehicle images that share the same target identity in the vehicle image library captured by surveillance cameras under the premise of a given search target(usually a vehicle image).Compared with license plate recognition,vehicle re-identification can overcome the limitations of monitoring scenes and shooting perspectives by mining vehicle identity information from vehicle visual features,which can greatly expand the applicable scenarios of vehicle search and improve its efficiency and intelligence.It plays a significant role in the construction of intelligent security systems and smart cities.The complex mapping relationship between vehicles’ visual features and their identities brings great challenge to vehicle re-id for complex enviroments,as well as that different re-id problems caused by different monitoring scenarios.To this end,this dissertation aims at three critical re-id tasks in complex scenarios: Disuigsed vehicle re-id,Near-duplicate vehicle re-id and Multi-vie vehicle re-id,addresses three challenging and valuable vehicle re-identification problems in complex scenes,namely,robust feature learning for anti-disguise vehicle re-id,fine-grained identification of similar vehicles,and stable representation and measurement of vehicle features from multiple viewpoints.The main research is shown as follows:(1)Anti-disguise vehicle re-id based on local structure-aware feature.In disguised vehicle re-id situation,addressing the difficulties in feature robust embedding caused the two common types of vehicle camouflage behavior: body color change and body pattern alteration,this dissertation studies the spatial distribution of the features in the front window area and proposes a re-id method based on the local structure-aware representation of vehicles.This method is oriented to the reidentification scene with camouflage behavior from the front view of the vehicle,encodes the spatial location and context information of the salient signs in the window area into the feature learning,and features the area from the semantic dimension and the structure-aware dimension of the front window area of the vehicle.It can greatly reduce the influence of disguise information vehicle re-id,improve the spatial perception ability of local features,and reduce the dependence of the feature learning model on data.In the next,using vehicle model label as supervision information,the feature learning network is applied to extract the vehicle model features that are less affected by camouflage at the global scale,and finally cooperate with the front window area features to complete the cross-scale anti-disguise vehicle feature representation to achieve high accuracy re-id results in the confrontation environment.On the self-built disguise vehicle re-identification dataset,the proposed method improves the Rank1 index of the best method by 6.9%,and the Rank5 index by 7.3%.This method can be applied in security checkpoints widely and has important application value in video surveillance.(2)Similar vehicle re-identification based on multi-granularity feature learning and enhancementIn near-duplicate vehicle re-id situation,to eliminate the challenges of telling similar vehicles apart,which caused by the limited visual differences between vehicles with the same/similar models and the complex variation between inter-class and intra-class of vehicles,this dissertation studies the joint learning and enhancement of multi-granularity vehicle features and proposes a multi-objective joint learning method with adaptive optimization stepwise,realizing efficient optimization of the distribution of samples in the feature embedding space.In addition,this method also designs a variety of hybrid attention modules to enhance the visual information of different visual scales and feature dimensions,and achieve accurate embedding of high-distinguishing visual features in multiple feature dimensions.Compared with the SOTA re-identification methods,our method achieves the highest Rank1 indicators on the three test sets of Vehicle ID: 81.1%,78.5% and 75.3%,which effectively improves the identification accuracy of similar vehicles.(3)Multi-view vehicle re-identification based on viewpoint awareness and selfattentionIn multi-view vehicle re-id situation,aiming at solving the huge changes of the overall vehicle visual features and the difficulty of local feature alignment caused by the diversity of viewpoint variation and the complexity of the shooting environment,the dissertation studies the change mechanism of the vehicle visible features under multiviewpoints and proposes a local feature similarity with viewpoints aware,achieving efficient feature alignment.At the same time,self-attention and dynamic self-attention modules based on Transformer and visual element position encoding are designed to learn long-range aware and space aware feature representations to realize stable vehicle feature representation and measurement with multiple viewpoints.Compared with several cutting-edge SOTA methods,our method achieves the highest m AP indicators on the three test sets of the multi-view vehicle re-identification dataset VERI-Wild:81.7%,75.9% and 68.0%,realizing multi-view high-precision vehicle re-identification.