Research On Maneuvering Obstacle Motion State Estimation Method For Intelligent Vehicle

In the last decade, driverless system has been regarded as the best solution of vehiclesafety and convenience of transportation. In this research field, multi-sensor perceptualsystem which has been used to help or take place of the driver’s perception of trafficenvironment is a basic task, essential condition and decisive factor of vehicle drivingassistance and unmanned driving systems‘operating reliability and accuracy.In this article, an intelligent vehicle platform based on multi-sensor system includinglaser radar, machine vision, GPS/INS and so on has been established. Based on this platform,this article has solved the problems of obstacle motion state estimation and multi-targetassociation, these research results can be used in radar-based and machine vision-basedobstacle estimation system directly.In this article, three main issues have been solved:1. In radar‘s measuring range, there are more than one obstacles, so distinguishing themfrom each other according to their respective source must be done, This is multi-targetassociation, which is the basis of obstacle motion state estimation.For realtime performance, rectangle measurement gate has been choose; In this article, animproved NNDA algorithm has been developed, obstacle size was added to build a differential function‘to decide the difference‘between measurements and obstacles; andweighting efficients was adjusted when distance between obstacle and radar was becominglarger and larger.2. Obstacles‘measurement which come from radar is mixed with noise; laser radarcannot measure obstacles‘velocity and acceleration directly, so the obstacle motion stateestimation algorithm which can decrease the noise and estimate velocity and accelerationmust be established. This article introduced Current Statistical Model‘from military field, which is using the modified Rayleigh distribution to describe acceleration. This model ismore aligned with the actual acceleration change laws of obstacles. This article alsoestablished adaptive Kalman filtering to estimate obstacles‘motion state. The motion stateestimation result is better than the results of Kalman Filtering based on CV and CA model;In this article, the motion modeling and filtering are based on absolute motion in absolutecoordinate, which can improve the accuracy by decreasing the influence from the movingmeasuring bases.3. Simulation and Real vehicle experiment validation: We implemented simulation andreal-car tests to validate algorithms of maneuvering target motion estimation and multi-targetassociation in typical condition, including constant speed, constant acceleration andtime-varying acceleration. The adaptive Kalman filtering based on CS Model canaccomplish accurate estimation task in all this conditions.When the distance betweenobstacle and radar was becoming larger and larger, the weighting efficients are adjustedaccordingly. In30real-car tests, the accuracy of multi-target association algorithm isapproximately90%.