Probability Hypothesis Density Filter Based On Adaptive Point Cloud Clustering For Multiple Vehicle Tracking

In decades,self-driving,the intersection of artificial intelligence and traditional vehicle engineering,has been widely concerned and hugely developed.Environmental perception technology is defined as the process that the intelligent vehicle system understands the surrounding traffic environment by processing the raw data of perception sensors.Accurate and comprehensive environmental perception,a research hotspot in self-driving,is the basis for the intelligent vehicle to achieve reasonable path planning and decision-making.Lidar is one of the most prevailing perception sensors in self-driving system.It can accurately obtain the distance information of surrounding targets,and has strong anti-interference ability.Moreover,its price is showing a dramatic decline.Vehicle is the most common traffic participant.The first and most important step of environmental perception of intelligent vehicle system is to detect surrounding vehicle targets and track their state accurately in real time.Considering these factors,we studies the vehicle detection and tracking based on 3D point cloud,and proposes a pipe of algorithms to process the raw data of point cloud and finally obtain the real-time state of the target vehicle.Firstly,the paper proposes a vehicle pose estimation algorithm based on adaptive 3D point cloud clustering.This algorithm contains three parts.The first part is ground points extraction algorithm based on ground state tracking using particle filter,which is used to extract ground points and delete them to get points belonging to vehicle targets.It avoids processing massive point cloud data in every frame like traditional algorithms,and obviously improves the real-time performance.The second part is to cluster remaining point cloud using adaptive threshold-based radially bounded nearest neighbor strategy.This algorithm improve the accuracy of clustering by adjusting distance threshold according to the characteristic of Lidar data.The final part is to estimate the vehicle target state(position,speed,heading angel)using likelihood-field model.This part of algorithm establishes the likelihood-field model of vehicle target according to measurement character and improves the accuracy by iterative optimization.In addition,the paper proposes the multiple vehicle tracking algorithm based on multi-model probability hypothesis density filter.the algorithm can effectively deal with the targets’ birth and death situation because of random set theory.At the same time,considering the diversity of vehicle motion in the actual traffic environment,the algorithm uses multiple-motion models to describe the target vehicle maneuvering,which can track the vehicle targets with high mobility accurately.