Multi-Extended Object Tracking with Generalized Physical Models

Abstract

The objective of sensor data processing is to extract the most informative knowledge possible by fusing prior information with imperfect sensor measurements. This dissertation focuses on model-based approaches for multiple extended object tracking that represent sensor and object characteristics as accurately as possible based on physical principles. <br/> In the first part, a novel approach for extended object tracking is presented. First, the measurement spread and center of gravity are estimated using a known approach from the literature. Subsequently, synthetic data are generated for a hypothetical estimate based on Virtual Measurement Models (VMMs), and the center of gravity and spread of these synthetic measurements are compared with those of the real measurements. Based on the resulting error, the test estimate is adaptively adjusted. The stability and convergence of this adaptation algorithm are proven in this dissertation. Virtual measurement models can represent arbitrary sensor and shape characteristics, enabling, for example, the accurate modeling of lidar sensors through a combination of contour and interior measurements. Assumptions that measurements are uniformly distributed over the contour or volume of an object are therefore not required. <br/> If it is unknown which shape best represents the object, or whether the measurements originate from the contour, the interior, or a combination of both, multiple virtual measurement models can be applied in parallel. The resulting artificial measurements can then be used for classification purposes. <br/> To evaluate the proposed VMM approach, extensive simulations are conducted and real recorded lidar data are analyzed, primarily originating from maritime scenarios. The VMM approach is particularly suitable for such environments because the objects encountered can be described very well by basic geometric shapes. For example, the shape of a motorboat resembles an ellipsoid, whereas a sailboat can be approximated effectively by an elliptical cone. <br/> The second part of this dissertation addresses the birth densities of tracking algorithms and the initialization required when an object is detected for the first time. Existing approaches place birth densities wherever measurements were observed in the previous time step. In scenarios with many false measurements, however, this leads to high computational costs and increases the risk of false detections. Therefore, the birth density is modeled here based on physical properties: the sensing range of a sensor is limited, and under unobstructed visibility, newly appearing objects must enter the surveillance area through its boundary; otherwise, they would already have been detected earlier and would therefore not be new. However, the exact location and width of this surveillance boundary must be determined adaptively. For this purpose, information about the initial positions of previously detected objects can be exploited. Since point symmetry can be assumed, particularly for lidar sensors with 360° coverage, a circular birth density in polar coordinates is introduced. The adaptive birth densities are evaluated extensively in simulation studies, and their relevance in real-world environments is demonstrated using a maritime scenario.