Research On Target Tracking Technology Based On Multi-source Sensor Information

Sensor target tracking refers to the continuous estimation of the quantity,state,and motion changes of targets based on noisy measurement data obtained from sensors.With the increasing demands for accuracy,real-time performance,and robustness in military and civilian fields such as maritime and aerial surveillance,autonomous driving,and mobile robotics,a single-source sensor is limited by its own physical parameters such as coverage range and resolution,resulting in limited target perception and tracking capabilities.Therefore,to compensate for the limitations of single-source sensors,it is of great research significance to integrate information data from multiple sources sensors and conduct collaborative research on multi-source sensor fusion for target tracking.In practical scenarios,the increasing variety of sensors and targets,unknown target quantities,increased maneuverability,enhanced environmental noise,dynamic variations,and other factors have made multi-source sensor target tracking face challenges such as difficulties in multimodal sensor modeling and fusion,mismatch between single target models and actual motion,the difficulty of balancing accuracy and timeliness in multi-target tracking,and dynamic modeling of multiple target types.In this dissertation,focusing on the above challenges,we delve into key technologies such as multimodal sensor data registration,complex motion target modeling,variational Bayesian and deep learning joint modeling,multi-target association,and interactive fusion filtering,aiming to improve the accuracy,robustness,and real-time performance of target association,tracking,and prediction.The research content and innovative work of this dissertation mainly include the following chapters:1.Chapter 2 focuses on the problems of significant differences in characteristics among multiple sensors such as radar,inertial measurement unit(IMU),Light Detection and Ranging(Lidar),and the difficulty of accurately modeling highly maneuverable targets.It investigates the algorithm for spatiotemporal registration of multiple sensors and establishes mathematical models for target motion.Specifically,for spatiotemporal registration,an improved virtual least squares algorithm is proposed to achieve time registration of heterogeneous sampled data from multiple sensors.A fast vector transformation method is designed for spatial registration,avoiding complex function computations and addressing the non-uniformity of data in time and space,thus improving the accuracy and reducing the complexity of spatiotemporal registration.To accurately describe highly maneuverable targets,a fusion of uniform velocity and uniform acceleration interactive multi-motion model set is constructed to adaptively track highly maneuverable targets in different dimensions.Simulation and experimental results validate the effectiveness of the proposed methods,laying a theoretical foundation for subsequent multi-sensor fusion and multi-target tracking.2.Chapter 3 addresses the challenges of accurate and timely multi-sensor multihighly maneuverable target association and tracking.From a system-level perspective,a comprehensive framework for multi-sensor fusion target tracking system and algorithm is established.The three-dimensional interactive multi-motion model set is expanded,and a point-track data distributed interconnection method is proposed.An adaptive separation of multi-target tracks is designed based on association criteria,and an improved interactive heterogeneous sensor Bayesian fusion filtering algorithm is proposed to fuse data from multiple sensors,achieving accurate and real-time tracking in complex multi-target motion scenes.Simulation and experimental results verify that the proposed methods,after fusing different types of radar information,improve the accuracy of association,tracking,and timeliness in chaotic and intersecting complex multi-object motion.3.Chapter 4 addresses the difficulties in mismatch between traditional sensor measurements and target motion models and actual measured data,as well as the challenges in solving the joint integration of posterior distributions from multiple sensors.A fusion method based on variational Bayesian and deep learning is proposed.By deriving the evidence lower bound(ELBO)optimization lower bound through variational Bayesian and latent state space modeling,a bidirectional normalization flow deep learning network is designed to learn the complex posterior distribution of targets through ELBO.The proposed method achieves accurate expression of target tracking and prediction posterior.In simulation experiments on nonlinear Lorenz force systems and real-world experiments on the KITTI dataset,after fusing information from different modalities of sensors such as global positioning system(GPS),IMU,and Lidar,the proposed method outperforms traditional model-based and popular data-based methods in terms of tracking and prediction accuracy,stability,and interpretability.4.Chapter 5 addresses the challenges of accurate association and tracking in complex scenarios with increasing target variety,dynamic changes in target quantities,and unknown clutter.A multi-sensor fusion interactive trajectory Poisson multi-Bernoulli algorithm is proposed.An interactive motion model is designed to provide a unified representation for different categories of targets and nonlinear target state changes.A cost matrix is constructed to transform the data association problem into an optimal assignment problem,and the Murty algorithm is introduced to reduce computational costs.The proposed method overcomes the limitations of existing random finite set-based tracking methods that only consider target number changes and do not consider target categories,achieving accurate tracking and prediction in complex scenarios with multiple and dense targets appearing and disappearing.Experiments on the KITTI target tracking dataset validate the improvement in tracking and prediction accuracy and stability in complex scenarios such as multiple vehicle movements and dense pedestrian environments after fusing information from different modalities of sensors such as Lidar,IMU and images.This dissertation provides a comprehensive framework and multiple innovative algorithms for target tracking by using multi-sensor fusion.These methods and algorithms address various challenges encountered in multi-source sensor target tracking.By designing multiple simulations and real-world experiments using the KITTI dataset,the proposed methods and algorithms have been validated to significantly enhance the accuracy,robustness,and real-time performance of target tracking and prediction under various complex factors.These research findings contribute to the advancement of target tracking in diverse applications by promoting their capabilities.