| Autonomous vehicles,as the next-generation transportation technology,hold immense potential in enhancing traffic efficiency and safety,as well as reducing energy consumption.As a central element in vehicular intelligence,the perception system must accurately recognize the static environment and dynamic targets around the vehicle to ensure safe and reasonable decision-making and control by the autonomous driving system.The primary challenge currently is autonomous driving in complex dynamic scenes,where the motion estimation of dynamic targets is a core task for the perception system.To extend the range of perception and improve the robustness and accuracy of motion estimation,multi-source sensors and information fusion offer an effective approach in designing perception systems for complex dynamic scenes.With the development of vehicle automation technology and the widespread adoption of vehicular network technology,motion target estimation with multi-source information fusion faces new challenges.On one hand,the perception system needs to integrate more sensors to meet the all-around perception requirements of upgraded vehicle automation functions,but the rapid iteration of vehicle automation systems requires the perception system to be able to compatibly deal with the problem of multilevel target information mixing brought by low-level perception subsystems and new sensors.On the other hand,vehicular network technology provides richer perception information for connected and automated vehicles(CAVs),improving the inherent limitations of single-vehicle sensors such as perception distance and occlusions.However,the complexity of information sources makes the correlation between information very complex,and dealing with the reference bases and asynchronous issues of information is difficult.Moreover,connected information depends on the current communication environment,such as effective communication distance affected by weather and obstructions,and the number of connected vehicles within the effective communication range.Therefore,how to improve the utilization rate of connected information also poses a new challenge.To Address these issues and challenges,this paper revolves around the technical route from individual vehicle intelligence to connected collaboration,based on the research of single-vehicle multi-sensor fusion method,progressively researches on the multi-vehicle information fusion method merging connected and localization information,as well as the cooperative perception method under communication-limited conditions:Firstly,to solve the multilevel target information mixing problem of advanced vehicle automation functions,a flexible and scalable hybrid multi-sensor fusion architecture is proposed.Combining centralized object-level fusion with track-level fusion,this architecture provides the scalability of new sensors and compatibility with low-level perception systems.Using covariance shift invariance,a multi-coordinate system centralised estimation algorithm is designed to integrate difficult-to-register nonlinear millimeter-wave radar measurement registration and estimation fusion steps,ensuring optimal centralized estimation performance.A global feedback multi-sensor data association algorithm is designed to decompose the whole multi-sensor association problem into a twostage bipartite graph matching problem by combining with the secondary track association,which solves the complex track association combination problem caused by complex sensor crossings and overlapping regions.Simulations and real-vehicle experiments effectively verify that this method can achieve stable all-around target tracking,as well as lowcomputation real-time performance and feasibility on the vehicular computing platform.Then,considering the integration of connected information to improve local perception,which leades to the complex information correlation,uncertainty,and asynchrony brought by connected information fusion,a multi-vehicle information fusion method under localization uncertainty is proposed.Split covariance intersection fusion algorithms for any number of sensors in two asynchronous cases are derived,and the correlation of localization and connected information is handled in stages by explicit modelling of information correlation.By the augumented state of localization and target motion estimation,the unification of coordinate bases and uncertainty fusion of locally fused intermediate quantities are achieved based on Gaussian-weighted integration.The boundedness analysis of the covariance of the entire multi-vehicle information fusion method is analyzed,ensuring the effectiveness of the algorithm in theory.Simulations and real driving trajectory data tests verify the improvement in motion estimation accuracy under different asynchronous scenarios and ensure safe consistent estimation of the method.Next,in view of the problem that the performance of multi-vehicle information fusion is difficult to be guaranteed in the communication-limited conditions of vehicular network,a distributed cooperative perception method under communication limitations is proposed.At the problem level,the communication topology of the vehicular network is modelled as an undirected graph by introducing a graph-theoretical modeling approach.At the methodology level,a centralized local fusion-distributed connected fusion cooperative perception architecture based on consensus algorithm is designed,which realizes networklevel information utilisation and information consistency among network nodes through the average consensus communication iterative protocol,solving the problem of insufficient available connected information under communication-limited conditions,while ensuring the consistency requirement of connected collaboration in environmental cognition.The optimal consistency information weight is theoretically analyzed and designed,ensuring the overall asymptotic optimality of the cooperative perception algorithm.The Lyapunov method is used to prove the stochastic stability of the estimation error.Simulations verify that this method significantly improves estimation accuracy under communication-limited conditions and is robust to time-varying communication topologies in actual applications.Finally,futher considering the maneuvering characteristics of road target motion under communication-limited conditions,a distributed multi-model cooperative perception method considering the uncertainty of motion patterns is proposed.At the problem level,the motion uncertainty of maneuvering targets is described as jumping Markov process with a finite model set by introducing the concept of the jumping Markov system.At the methodology level,a model probability update method that simultaneously utilizes local sensor and network information is derived,for which a distributed solution method is designed based on the consensus algorithm.By combining this model probability solution method with multiple parallel mode-conditional filters,the distributed cooperative perception algorithm is extended to the interactive multiple model framework.Simulations and complex traffic scene tests based on the SCANe R software verify that this method,while ensuring estimation accuracy under communication-limited conditions,achieves accurate estimation of the motion uncertainty and state of maneuvering targets.In summary,this paper presents a comprehensive and in-depth study of multi-source information fusion techniques for motion target estimation,from individual vehicle intelligence to connected collaboration environments,addressing the challenges of environment perception poesd by the complexity and dynamics of vehicle driving environments.The effectiveness of the proposed methods is thoroughly verified through extensive simulations and real-vehicle experiments.This research provides a solid technological foundation and theoretical support for the design of perception systems in CAVs,playing a significant role in advancing the development of safer and more efficient autonomous driving technologies. |
PDF Full Text Request