Vehicle Re-Identification And Trajectory Prediction In Intelligent Vehicle-Infrastructure Cooperative System

The number of road vehicles has risen sharply in past decades since the rapid development of social economy and living standards,as well as the improvement of road infrastructure,bringing a variety of problems,including the most important road safety.Cooperative vehicle-infrastructure system(CVIS)is attracting more and more attention according to its potential for reducing both congestion and accidents.Unfortunately,at the current level of technology,the perception capability of a single smart car is limited by the obstacles,field of view and performance of sensors.However,in the intelligent vehicle-infrastructure cooperative system(IVICS),the smart cars can communicate with the roadside units and other adjacent smart cars to interchange information and fuse detections,which will effectively make up for the perception limitation,thereby improving road safety.This thesis studies vehicle re-identification and vehicle trajectory prediction,two of the key technologies in IVICS,and proposes one deep-learning-based method on each aspect respectively.This thesis is composed of two main parts with these contributions:(i)A large number of existing methods are reviewed and discussed,the difficulties and challenges encountered in vehicle re-identification and vehicle trajectory prediction are analyzed respectively,ideas and possible solutions are summarized then.(ii)To address the problems caused by viewpoint variant including intra-class differences and inter-class similarities in vehicle re-identification,this thesis divides vehicles into several different but practical components,extracts and aligns features from different components to avoid the problem of intra-class differences.Then,the importance of each component is dynamically weighted using attention mechanism,making the network learned discriminable feature extraction against inter-class similarities.To further improve representative power of adopted attention mechanism,this thesis introduces a novel attention fusion mechanism,summarizing two different types of attention weights according to the area occupied by and features contained in different components.(iii)To address the interaction between vehicles in trajectory prediction,this thesis adopts graph neural networks,describing the vehicles in specific scenarios as nodes and their connections as edge in a graph,instead of traditional methods to model the spatiotemporal interaction between target vehicle and adjacent vehicles.To obtain a lower deviation at long horizon,an innovative method applies maneuver is proposed.Firstly,interactions between vehicles are first encoded to estimate maneuver.Secondly,together with encoded interactions,estimated maneuvers are embedded and feed into decoder.Finally,predicted horizons of target vehicle are generated by decoder under the supervision of maneuvers.(iv)To verify effectiveness of proposed methods,varieties of experiments are conducted using diverse public large-scale datasets on each aspect respectively.The results on popular vehicle re-identification dataset Ve Ri shows proposed baseline without attention fusion mechanism has already outperformed with a large margin between previous methods.Furthermore,proposed baseline with attention fusion mechanism achieves the performance of m AP at 80.5%,better than applying each attention mechanism respectively.The results on popular vehicle trajectory prediction dataset high D shows proposed vehicle trajectory prediction method also achieves state of the arts performance.Both proposed vehicle reidentification and trajectory prediction method have broad application prospects in IVICS.