Subject Area
Mathematics
Description
This project addresses the challenge of state estimation in real-world conditions by fusing data from multiple sensors using the Kalman Filter. To ensure numerical stability, we use the Joseph Form covariance update, which guarantees valid results but introduces significant computational overhead due to its complexity. To overcome this limitation, we implement a parallelized solution using custom CUDA kernels on a GPU, distributing matrix operations across thousands of threads rather than relying on sequential CPU execution. Through systematic benchmarking across matrix sizes ranging from 4×4 to 4096×4096, we identify a crossover region where GPU performance surpasses CPU efficiency. This work shows that a numerically robust Kalman Filter can be executed in real time on edge GPU hardware, enabling more reliable autonomous systems.
Publisher
Providence College
Date
Spring 4-22-2026
Start Date
4-22-2026 3:00 PM
Type
Poster
Format
Text
.pdf (text under image)
Language
English
