Research On Multi-Sensor Fusion Target Based On Information Matrix Tracking Method

Intelligent driving technology has been widely used in the field of automotive transportation,especially in reducing traffic accidents and alleviating traffic congestion,and has now become a major research hotspot in the automotive industry.In the study of intelligent driving technology,the ability to maintain stable trajectory tracking of intrusion targets in the in-vehicle sensor network and accurately fuse information from different sensors is a prerequisite for intelligent vehicles to make correct judgments,and is an important guarantee for the safe travel of intelligent vehicles.However,the processing of in-vehicle multi-sensor information fusion contains the following two difficulties: one is the accurate estimation of the motion state of the detected target and the stable tracking of the detected target;the other is the effective correlation and fusion of the targets tracked by different sensors to obtain a robust estimation of the actual target.Thesis focuses on the multi-source information fusion method,which mainly includes the state estimation method of non-linear motion target tracking based on multisource information and the correlation and information fusion method between the information of the same target tracked by multi-source sensors.The main research elements of thesis are as follows:1.For sensor target tracking models and data association problems,thesis introduces several different types of manoeuvring target motion and derives several types of commonly used target motion models for smart cars.Given that the state estimation process of most in-vehicle sensors for intrusion target tracking is a non-linear state estimation problem,thesis introduces the Kalman filter principle.In light of the practical situation,and given that non-linear optimal filtering is difficult to implement,thesis provides a further introduction to extended Kalman filtering and traceless Kalman filtering.2.A multi-sensor target tracking and detection algorithm based on information matrix fusion is proposed to address the problems of low accuracy,inaccurate parameters and scattering of filtering algorithms in existing information fusion algorithms.In this method,an asynchronous sensor-to-global fusion strategy is used to immediately process the sensor data in a distributed fusion structure to provide the driver with up-to-date information about the vehicle.Optimisation of the filter divergence problem also allows for improved accuracy of the overall data fusion system and enhanced safety and comfort of the autonomous driving.3.To address the problem of limited applicability of the algorithm to the scenario,an information matrix fusion algorithm combined with neural networks is proposed to obtain an information fusion algorithm with better accuracy and more adaptive capability.The algorithm exploits the similarity between the neuron structure of Kalman filtering algorithm and Recurrent Neural Network(RNN)network,and by comparison,it is found that Back Propagation Through Time(BPTT)can also be trained in combination with the information matrix fusion algorithm.By embedding a trainable structure for the information matrix fusion algorithm,more robust filtering results are obtained.