The Research On Context-enhanced Multi-target Tracking Methods For Intelligent And Connected Vehicle

The technology of intelligent vehicle is important for the development of future automobiles.It involves environment modeling,multi-sensor fusion perception,situation estimation and cognition,decision-making,control,and other alternatives.It is an interdisciplinary field of artificial intelligence,control,communication,and computer.As one of the cores and key technologies in intelligent vehicle technology,environment perception is the focus and hotspot of current research and the main reason of restricting the development of intelligent vehicles.Due to the defects of sensors’ properties,traditional multi-sensor perception technology cannot satisfy the needs of road perception in mountainous areas and other complex environments.Therefore,taking context information as a new entry point for environment perception technology research,a novel method of vehicle motion model and multi-sensor fusion considering context information is proposed.Incorporating with connected information,the connected cooperative tracking method and system are proposed.The novel target tracking system has obvious theoretical significance and application value for improving the real-time accuracy and reliability of on-board target tracking.The main content of this dissertation has the following aspects:(1)The motion model construction of low maneuvering targets in the target tracking problem is studied.A vehicle tracking model based on a microscopic traffic force field model is proposed for low maneuvering vehicles.The method considers the traffic scene and road constraints in the two-dimensional road coordinate system and constructs a microscopic traffic potential field model.Then,based on the differences in vehicle longitudinal and lateral behaviors,the lateral and longitudinal motion models are proposed respectively,and the interacting multiple model is used to estimate various motion models.At the same time,considering that the transition probability matrix in the traditional interacting multiple model is usually determined by prior knowledge,a quasiBayesian recursive algorithm with an adaptive transition probability matrix is proposed to deal with the switch between the lateral maneuver models,which implements the rapid and accurate transfer of the lateral motion model and improves the target tracking accuracy.Finally,the tracking frameworks of point target and the extended target are established according to the different characteristics of sensors,and a complete tracking system is built to track the vehicle targets.(2)The motion model construction of high maneuvering targets in the target tracking problem is studied.A semi-Markov model tracking method based on context information is proposed for high maneuvering vehicles.Considering the high maneuverability of the targets,a semi-Markov model based on sojourn time is constructed to model the longitudinal motion of the target,and the state probability matrix among the models is calculated adaptively from the accumulated sojourn time.This dissertation analyzes the cause of high maneuvering behavior considering the external context information and constructs the Bayesian network associated with the context to estimate the sojourn time of each model in the semi-Markov model.To model the lateral movement of vehicle targets,the road center point is used to construct a pseudo-measurement to track the lateral movement with constraints.The proposed method can effectively reduce the uncertainty of estimation and deal with the transformation relationship between models more accurately to improve the accuracy of tracking high maneuvering targets.(3)The multi-sensor fusion and tracking problem under the framework of random finite set is studied.An uncertain multi-sensor fusion and tracking method based on a random finite set is proposed for multi-source heterogeneous sensors(camera,radar,and Li DAR(Light Detection and Ranging)).The unified target state kinematics and measurement model of heterogeneous sensors is constructed,and the kinematics model and target recognition attributes are modeled in a unified way.A framework of multitarget joint detection and tracking recognition of heterogeneous sensors is proposed.Then,the uncertainty measurement model under the framework of a random finite set is established for the fuzzy measurement of heterogeneous sensors.On this basis,combining with the different field of view and uncertainty scene influence on the sensor’s measurement,a multiple sensor fuzzy fusion method considering context is proposed.The proposed method can ensure that the on-board sensors can perceive the surrounding targets reliably and stably in the complex environment,maximize the use of the optimal fusion measurement information,and eliminate the adverse impact of individual sensors on the whole fusion system.Poisson Multi-Bernoulli Mixture method under the framework of random finite set is used to track the target,and the joint detection,recognition and fusion tracking of multiple-targets multiple-heterogeneous sensors are implemented.(4)The target tracking method using connected information for intelligent and connected vehicle is studied.Aiming at different kinds of network auxiliary information,a target tracking system based on vehicle to vehicle(V2V),vehicle to infrastructure(V2I)and vehicle to network(V2N)cooperative information is proposed for intelligent and connected vehicle.In the V2 V cooperative tracking,a target tracking system based on radar and connected information fusion is proposed.By using the identity and position information sent by V2 V communication,the system can improve the accuracy and reduce the computation of data association algorithm.Aiming at the tracking problem of extended target detected by Li DAR,a target tracking system using the extended cooperative information fusion is constructed in this dissertation.The system uses the shape and heading information sent by the V2 V to construct a double-tracking gate to reduce the clutter of Li DAR measurement and improve the tracking performance.Aiming at the tracking problem of V2 I,three classical track fusion methods are analyzed and verified by experiment,and the optimal fusion method is selected to construct the track fusion system of roadside unit and on-board sensors.Therefore,the on-board perception is extended to connected cooperative perception,which increases the perception range and improves the tracking accuracy and reliability.