Research On Key Technology Of Lidar Based Object Detection And Tracking For Intelligent Vehicles

Intelligent Vehicle (IV) is a major component of Intelligent TransportationSystems (ITS). Environment perception plays the role of cornerstone in the IVarchitecture. It provides necessary information for the control of the vehicle. Thus theperformance of environment perception system affects directly on the driving safetyof the intelligent vehicles.In this dissertation, laser radars were adopted to acquire the range information ofthe environment. Aiming at the key issues in dynamic environment perception system,i. e. obstacle detection and mobile object detection, the following aspects were mainlystudied in this dissertation:(1) Vehicle detection in the structured road environment was discussed byfeature based method. According to the trait of the range information reflected fromvehicles, an adaptive cluster algorithm based on point distance was proposed. Thedistance threshold could be adjusted according to the relative orientation and distanceof the objects and ego vehicles. To match the shape of a vehicle, the IEPF algorithmand least square method were adopted within every cluster to fit a straight line orpolygonal lines. Feature vectors were abstracted from the fitted shapes. Experimentresults showed thatthe cluster accuracy was improved by the algorithm proposed inthis dissertation in comparison of by a fixed distance threshold. The improved dectionprovides a good basis for vehicle tracking.(2) Considering the complexity of the city road where many kinds of static anddynamic objects are included, occupancy grid maps were built based on Bayesianframework and evidence theory separately. Aiming at the problems in the polarcoordinate based inverse sensor model, an inverse sensor model based on Cartesiancoordinate was built and used to translate the range information into a localoccupancy grid map. Quadratic interpolating in the fusion of the local map and theglobal map was avoided when this inverse sensor model was used. And thus, thetranslation accuracy increased and the calculation amount decreased.(3) Mobile object detection methods were proposed according to the differentuncertainty reasoning algorithms in Bayesian framework and evidence theory. Thesemethods were comparably analyzed by both simulation and experiments. Simulationand experiment results showed that the measurement noise was effectively filtered out when the occupancy grid map was updated by the original Bayesian inference orDempster’s rule of combination, and the amount of false alarms was also reduced. Bycomparing the inconsistency of the MG at time step t-1and the SG at time step t,mobile objects were clearly and wholly detected in the occupancy grid map updatedby these two reasoning algorithms. However, the measurement noise was not able tobe filtered out, and the detected mobile objects were incomplete when the occupancygrid map was updated by PCR2in DSmT framework or the corrected Bayesianinference. Mobile objects were abstracted from the occupancy grids by digital imageprocessing technologies, and they were classified by fuzzy logic according to theobject width.(4) Mobile object tracking was carried out after detection. Considering theproblems in the application of deterministic data association algorithm in themulti-object-tracking of autonomous driving system, improvements wereimplemented to make it able to associate the uncertain number of objects, and toincrease the association accuracy. The improvement includes mainly two aspects:1)local optimization was used instead of global optimization to fulfill the practicalsituation that the number of moving objects changes. With the assistance of propertracker management policy, tracking of the uncertain number of objects could beachieved bythe improved data association algorithm.(5) In the simulation of multi-object-tracking in occupancy grid map,classification and kinematic attributes were both considered in the cost function inorder to increase the association accuracy. The classification attribute was fused andupdated by PCR2of DSmT, and the classification deviation between newmeasurement and tracked object was calculated through pignistic probability.Simulation results showed that the association accuracy was greatly improved whenthe classification attribute in data association was considered. It also suggests that thefusion result by PCR2is reliable in high conflict situation.(6) In the vehicle tracking experiment, a tracker management policy wasproposed.The tracking for the changing number of objects was achieved incombination of our proposed policy and the improved deterministic data algorithm.The four typical conditions: continuous tracking for two vehicles, the decrease in thenumber of tracked objects, the increase in the number of tracked objects andtemporary occlusion were studied in the experimental results. The tracking result bythe algorithm proposed in this dissertation coincided all with real situation recorded by the camera under the four typical conditions, which verified the effectiveness ofthe tracking algorithm proposed in this dissertation.