The recent advancement of the autonomous vehicle has raised the need for reliable environmental perception. This is evident, as an autonomous vehicle has to perceive and interpret its local environment in order to execute reactive and predictive control action. Object Tracking is an integral part of vehicle perception, as it enables the vehicle to estimate surrounding objects trajectories to achieve dynamic motion planning. The 3D LIDAR has been widely used in object tracking research since the mechanically compact sensor provides rich, far-reaching and real-time data of spatial information around the vehicle. On the other hand, the development of autonomous driving is heading toward its use in the urban-driving situation. In an urban situation, a robust detection and tracking algorithm is required due to increasing number of Vulnerable Road User (e.g. pedestrian and cyclist), heterogeneous terrain, inherent measurement uncertainties and limited sensor reach.

This thesis presents an integrated framework of multi-target
object detection and tracking using 3D LIDAR geared toward urban use. The framework combines occlusion-aware detection methods, probabilistic adaptive filtering and computationally efficient heuristics logic-based filtering to handle uncertainties arising from sensing limitation of 3D LIDAR and complexity of the target object movement. The implemented framework takes a raw 3D LIDAR data as input to perform multi-target object detection while simultaneously maintaining track of the detected objects' kinematic states and dimension in robust, causal, and real-time manner.

Robust detection is enabled by slope-based ground removal and L-shape fitting to reliably enclose bounding box to objects of interest in the presence of sensor occlusion. The tracker utilises three combined Bayesian filters (IMM-UK-JPDAF) which simultaneously tackle association uncertainties, motion uncertainties and estimate non-linear stochastic motion model in real-time. Logic-based rule filters are also designed to augment the rest of detection and tracking based on the understanding of LIDAR sensor limitation and occlusion characteristic.

The evaluation results using real-world pre-recorded 3D LIDAR data show the proposed framework can achieve promising real-time tracking performance in urban situations. Diverse datasets are deliberately chosen to evaluate if the MOT system is capable of working in a varying driving scenario. The benchmark results highlight that the designed and implemented MOT system is performing on par with the state-of-art works and yield satisfying accuracy and precision in most given urban settings.

One of the most crucial links to building an autonomous system is the task of decision making. The ability of a vehicle to make robust decisions on its own by predicting and assessing future consequences is what makes it intelligent. This task of decision making becomes even more complex due to the fact that the real world is uncertain, continuous and vehicles interact with each other. Sensors that perceive the real world and measure quantities such as position and speed of other traffic objects are inherently noisy and are further susceptible to external conditions. On the other hand, the road users’ intentions are stochastic and not measurable, and the presence of partial or complete vision based occlusions can make any measurements obtained useless. The decision making unit thus has to be aware of these issues and use the limited knowledge available to anticipate future situations that could unfold in an infinite number of ways to then maximize a reward (or minimize a cost).

In this thesis, a method to make automated longitudinal decisions along a predetermined path for autonomous vehicles in unsignalized urban scenarios is proposed. The decision mak- ing problem is formulated as a continuous Partially Observable Markov Decision Process (POMDP) with a discrete Bayesian Network estimating the behavior of other traffic objects. The future evolution of states are predicted using a multi-model Trajectory Estimator along with the digital map of the current driving area. Since continuous spaces make the belief space infinitely large, calculation of the value function for this problem becomes computa- tionally intractable. The presented solver algorithm approximates the value function instead of computing it directly, and additional optimizations reduce the belief space exploration area in order to improve performance. The results show that this single generalized framework is able to handle all the tested urban scenarios with a good safety margin in scenes with multiple traffic objects, even under zero visibility of the other traffic objects due to the presence of occlusions.