Dynamic Target Perception Method For Complex Driving Scene Based On Information Fusion

In the era of intelligence,unmanned driving technology is rapidly maturing with the rapid advancement of information science,and it is foreseeable that it will have a profound impact in many fields related to the national economy and people’s livelihood.In addition,to ensure autonomous and safe passage in complex environments,unmanned vehicles must have the ability to accurately perceive the driving environment and understand the driving situation.One of the most challenging tasks is dynamic target perception in complex driving scenes.This dissertation focuses on the task of dynamic target perception in the environment perception system of unmanned vehicles.Taking the actual complex driving scenes as the engineering practice background,it explores the application of information fusion technology in improving the ability of dynamic target perception.Although a lot of research work has been carried out in the field of dynamic object perception,the stability and accuracy of dynamic object perception cannot always be guaranteed in some challenging scenes.Therefore,this research project is proposed to solve this problem.And it takes challenging complex driving scenes as the point of penetration,analyzes the limitations of current methods in complex scenes,proposes targeted solutions,and conducts relevant experimental verification.Specifically,the major innovations and research achievements of this dissertation are listed as follows:(1)In terms of target detection and environmental element extraction,this dissertation first proposes a general ground segmentation method for multi-LiDAR fusion systems and a point cloud clustering method based on iterative rules.Secondly,this dissertation gives the connection and difference between the drivable area,the reachable area,and the road area in the unstructured environment,and clarifies the connotation of the drivable area.On this basis,two drivable area detection methods are proposed to adapt to different unstructured scenes,and they are based on Bayesian fusion and topological maps,respectively.Finally,this dissertation designs an interactive coupling framework for dynamic target tracking and drivable area detection,which can be applied to the interactive passage between unmanned vehicles and other dynamic targets in unstructured dynamic scenes.(2)In terms of data association,this dissertation proposes a feature dissimilarity measurement method for data association in complex scenes,which is utilized for the fusion of features with different attributes,and gives the physical meaning of the above fusion method using information entropy theory.Meanwhile,aiming at the problem of target loss and re-association in the tracking process,this dissertation designs a hierarchical association method based on spatiotemporal information and determines the priority of the association through the target uncertainty.In addition,aiming at the problems that the UAV camera and the UGV LiDAR cannot carry out accurate spatial calibration and traditional distance and appearance features fail in air-ground target association missions,this dissertation designs the Position-Ordering graph features for the cross-domain association to represent the relative position relationship between targets,and combined it with the above feature fusion method to achieve air-ground target level information fusion.(3)In terms of target pose estimation and tracking,this dissertation proposes a dynamic vehicle pose estimation and tracking method fused with motion information.To eliminate the observation error caused by target occlusion,the proposed method uses the motion information of the target as feedback to assist in target pose estimation.Meanwhile,a new heading normalization vehicle observation model is designed,and the matched filter and its cost function are constructed by taking the LiDAR measurement points as a two-dimensional Gaussian signal distributed on the model.On this basis,the nonlinear optimization method is used to obtain the optimal pose estimation of targets.Furthermore,the vehicle pose estimation and tracking method utilizes an interactive multitude model(IMM)tracker to capture the motion pattern of the target,and it combines IMM with a non-maximum suppression vehicle size estimation strategy,so that the method can be applied to vehicle targets of different sizes,and can achieve accurate target pose estimation and tracking without complete target measurements.