Detection And Tracking Of Vehicles Based On Vehicular 3D LIDAR Data

The Autonomous Land Vehicle(ALV)is one of the most important components for the intelligent transportation system.It has broad application prospect in fields of improving traffic and providing convenient transportation.The basis of the autonomous driving is the precise environment perception,especially for the complex urban environment.It is an essential requirement to accurately detect and track other traffic participants i.e.vehicles and pedestrian,for the environment perception system.This thesis mainly aims at the requirement for autonomous driving of intelligent vehicle in complex urban environment,and it carries out the research on detecting and tracking the vehicle which is one of the main traffic participants based on high-precision three-dimension LIDAR.The main results and contribution of this thesis are as follows:Firstly,because traditional segmentation algorithms usually tend to produce under-segmentation and over-segmentation errors,a novel method which adopts the sliding window models based on the view angles is proposed to propose vehicle-like regions fast.Combining the sliding window models in different view angles with LIDAR data,this method extracts representative weak feature to filter the impossible vehicle objects.This method overcomes under-segmentation and over-segmentation errors caused by traditional segmentation algorithms and object appearance variance caused by the changes of the observed distance and view angles.Further,this method can estimate the rough pose of the vehicles to initialize the state of vehicles for tracking.Secondly,we label some samples of vehicle on our LIDAR data and classify the samples with different distance and view angles according to the data distribution of three-dimension LIDAR.Then we use labeled samples to train simple classifiers and refine the open database.We establish new database of vehicles with our labeled samples and refined open database samples.Two new features about max height and points number along length direction of vehicle are extracted according to the characteristics of the vehicles.Combining the new features with common statistics features,the AdaBoost classifiers are trained to recognize the vehicles with different distance and view angles.These classifiers get satisfying results.Thirdly,an improved vehicle observation model based on point clouds probability integral is proposed to track the vehicle combining with particle filter.This method can estimate the pose and velocity precisely and deal with the tracking problems of vehicles which are occluded by other obstacles.This method gets an satisfying tracking results in crowded scenes.