Simulation-Based Data Fusion and Tracking
by Fabian Sigges
Date of Examination:2024-08-15
Date of issue:2025-05-15
Advisor:Prof. Dr. Marcus Baum
Referee:Prof. Dr. Marcus Baum
Referee:Prof. Dr. Benjamin Noack
Persistent Address:
http://dx.doi.org/10.53846/goediss-11253
Files in this item
Name:Thesis_Sigges_Revised.pdf
Size:1.29Mb
Format:PDF
The following license files are associated with this item:
Abstract
English
Autonomous systems ranging from self-driving cars and driver assistance systems, over autonomous vacuum cleaners to unmanned aerial vehicles rely on accurate perception of their surroundings to navigate and interact with their environment. To achieve fully autonomous systems a whole variety of problems has to be solved. Two of these problems are multi-object tracking (MOT) and Simultaneous Localization and Mapping (SLAM). MOT involves dynamically monitoring the movement of multiple objects concurrently and keeping track of their trajectories. SLAM on the other hand is concerned with building a map of the surroundings using mainly static objects as landmarks and features, while also localizing yourself in this map. In this work we aim at solving the two problems leveraging simulation-based approaches. Approximate Bayesian Computation (ABC) and Ensemble Kalman filter (EnKF) are simulation-based algorithms that appear to offer certain advantages over traditional algorithms like flexibility regarding nonlinear models. Both algorithms have however only gained little attention in MOT or SLAM literature. In MOT, employment of particle filters can be difficult due to complex or intractable likelihood functions. ABC attempts to circumvent this problem by a forward simulation step where only a generative function for measurements is necessary, but no costly computation of the likelihood function. Regarding SLAM, one of the major challenges is the size of the state vector and its associated covariance matrix. For a growing number of landmarks in the state vector handling of the covariance matrix becomes increasingly difficult. With its origin in numerical weather prediction one of the main properties of the EnKF is its ability to estimate very large state vectors with up to a million dimensions and as such, it appears to be a promising candidate to solve the SLAM problem. The main question throughout this work is, how well do these more general simulation-based approaches perform against traditional problem-tailored algorithms? To answer this question, the algorithms are formulated and adapted to fit the respective problem and then compared in common scenarios against established algorithms. The results show that in very general scenarios there is no real advantage for the simulation-based approaches. However, in corner cases, where traditional algorithms cannot be applied or require significant approximations or reformulations, simulation-based approaches work comparably well with very little effort.
Keywords: Multi-Object Tracking; Kalman Filter; SLAM; Particle Filter; Bayesian Filter
