| dc.description.abstracteng | Extended object tracking is an emerging research topic that is motivated by the rapid development of modern sensors. The traditional object tracking assumes a tracked object is far away from the sensor. Therefore, an object takes only one resolution cell and can be simplified as a point. However, due to the employment of near-field and high-resolution sensors, it is common for an object to occupy several resolution cells, and its extent is not negligible in many modern applications such as autonomous driving, robotics, and surveillance. Extended object tracking estimates both the kinematic state and spatial extension of an object based on a varying and unknown number of measurements. In this thesis, the object extensions are described as elliptical shapes. This thesis is devoted to three problems in the context of extended object tracking and has made three contributions respectively:
Evaluation metric
Between two ellipses that describe the same object, which one is better? Many elliptical extended object trackers have been developed, but no consensus exists on the measures for performance comparison. The Euclidean distance, which evaluates the location error for point object trackers, incorporates no shape error. Finding a simple and intuitive measure that combines both location and shape errors is not straightforward. Through the discussion and evaluation of the possible performance measures, the first contribution of this thesis is the proposal of using the Gaussian Wasserstein distance for evaluating elliptical extended object trackers.
Shape estimation
Given a set of measurements originated from one extended object, how to derive the kinematics and the shape of the underlying object? The estimation of object extension is challenging as it is a high-dimensional and non-linear estimation problem. The state-of-the-art elliptical trackers approximate the object shape as a symmetric positive definite random matrix, which couples the orientation and axes lengths. However, modeling the dynamics of orientation and axes lengths individually is useful for many applications. Therefore, the second contribution of this thesis is a single elliptical extended object tracker that explicitly estimates object kinematic state, orientation and semi-axes lengths. A closed-form solution is derived in the framework of recursive Kalman filter. Using the Gaussian Wasserstein distance as a metric, simulation results have shown that the proposed tracker facilitates the dynamic modeling of extended objects and outperforms the previous work on this topic.
Multiple extended object tracking
Knowing a set of measurements from multiple objects, what are the location and the shape of each object? The key to solve this problem is data association, i.e., determining the origin of each measurement. Many multiple extended object trackers rely on clustering techniques to obtain measurement partitions so that measurements generated from the same object are in one cell. Then, the measurement cells are assigned to potential objects using data association methods in traditional object tracking. However, the clustering process normally incorporates predicted object density heuristically and has high complexity. The third contribution of this thesis is a new multiple extended object tracker that employs an efficient measurement-object assignment method and using the single extended object tracker for shape estimation in contribution two. The new data association method calculates the marginal association probabilities by considering all measurement-object mappings, yet requires no clustering or explicit enumeration of assignments. The proposed tracker is tested using simulation and real lidar data. Results showed that the proposed tracker is more efficient and performs better than clustering-based trackers. | de |