Date of Award
Doctor of Philosophy in Electrical Engineering
Electrical, Computer, and Biomedical Engineering
In this dissertation, we investigated a multi-target tracking filter, specifically the Probability Hypothesis Density (PHD) filter, for tracking non-homogeneous targets with varying characteristics. In traditional tracking systems, targets are assumed to have fixed characteristics with constant sensor detection rates. However, targets like cyanobacterial blooms exhibit different sizes and shapes at each tracking step, introducing additional uncertainty for the tracking system.
In this study, we propose a novel formulation of the multi-target tracking problem, in which the target represents a variable shape and size subset of the state space. Due to the lack of a proven model and the non-linear nature of the system, we present a solution based on the particle implementation of the PHD filter. Simulations demonstrate promising results, allowing for a significant reduction in errors compared to measurements. Furthermore, we propose an optimization method for tailoring the proposed filter to the lake it is being deployed in. We employ a popular genetic algorithm to test combinations of variables and find the optimal values. The proposed method is tested on a satellite-based dataset collected over Lake Sabbatus, Maine.
Additionally, the thesis proposes a modification to the Gaussian mixture Probability Hypothesis Density (GM-PHD) filter to compute online the probability of detection (PD) and probability of survival (PS) for individual targets, eliminating the need for predetermined and constant PD and PS values. We introduce novel parameters to the filter and propose parameter updating schemes in the time update and measurement update steps. Experimental validation of the proposed filter is conducted through in-lab and real-world scenarios, including unmanned ground robots and Automated Surface Vehicles (ASVs) in fresh water lakes with real-time boat traffic. These findings highlight the potential benefits of the pro-posed filter in improving target tracking performance in complex environments.
Moreover, building upon the previous research, we extend the proposed filter to estimate PD and PS in a multi-sensor setup with different Fields of View (FOV). The filter is tested in simulation, in-lab with real robots, and in real-world using anunmanned surface vehicle attached with a camera sensor and a Lidar sensor. The results demonstrate the effectiveness of the proposed multi-sensor tracking system in accurately estimating target states.
Perera, Rupasinghe, "AUTOMATED SOLUTIONS FOR DYNAMIC SUBSET TARGET IDENTIFICATION AND TRACKING" (2023). Open Access Dissertations. Paper 1615.