Neural Compression for Multimodal Vehicle Data - Investigating Different Architectures for Neural Compression Models for Multimodal Time Series Data

Abstract

With the growing amount of data generated by modern vehicles, the need for edge efficient data processing and compression strategies becomes more apparent. As the amount of data grows so does the potential for machine learning models to extract insightful information from this data at later stages. Achieving high compression ratios while retaining only the essential components of the original data therefore becomes a primary goal of many compression strategies. Currently, the existing solutions for such a task on a resource-constrained device such as a vehicle are limited and often come with tradeoffs. This thesis investigates compression algorithms based on neural networks (commonly referred to as neural compression) which have shown promising results in other fields, such as image processing. Four neural codec architectures are explored and tested on industry-relevant datasets with compression ratio, data reconstruction quality and downstream utility retention as the main metrics investigated.