Researchers from the University of Cambridge and the University of Essex have unveiled a novel approach to enhancing simultaneous localisation and mapping (SLAM) systems by introducing FTI-SLAM, a federated learning adaptation tailored for thermal-inertial SLAM frameworks. This innovative system aims to tackle prevalent challenges associated with data privacy, high communication costs, and the need for generalisation while maintaining performance akin to traditional centralised SLAM models.
SLAM technology is critical for the navigation and mapping of autonomous systems, allowing them to ascertain their position and movement trajectory within unknown environments. This capability becomes particularly challenging in visually compromised settings, such as smoke-filled or dimly lit areas, where standard visual sensors struggle. The integration of thermal imaging into SLAM offers a solution, yet significant concerns regarding the transmission of sensitive visual data for centralised model training persist.
Haochen Liu and Hantao Zhong from the Department of Computer Science at the University of Cambridge, along with Weiyong Si from the School of Computer Science and Electronic Engineering at the University of Essex, collaborated to implement FTI-SLAM. Speaking about their study, the researchers highlight its feasibility in maintaining system performance while addressing pressing privacy and communication issues. This project utilises Flower, a comprehensive federated learning framework, to facilitate local model training, utilising various aggregation strategies.
The advantages of FTI-SLAM over traditional SLAM systems are threefold. Firstly, it fortifies data privacy by leveraging federated learning techniques, thereby preventing the need for transmitting sensitive thermal images to a central server. Secondly, it significantly lessens communication expenses by only sending aggregated model parameters instead of raw sensitive data. Lastly, FTI-SLAM improves generalisation through an inclusive training process involving diverse data sources, leading to more robust models that can be fine-tuned for specific user requirements.
The team tested FTI-SLAM using datasets originally collected for TI-SLAM, with experimental outcomes demonstrating that it could sustain comparable performance levels to the traditional model, even with limited computational resources and identical training data.
However, the current implementation of FTI-SLAM has been confined to six clients due to computational limitations and the need for additional data for simulations on a larger scale. Future developments may involve scaling the framework to accommodate a greater number of clients, enabling extensive simulations to assess the system’s performance under various configurations and the influences of data diversity.
Another critical aspect to address is the formulation of robust aggregation algorithms to guard against malicious attacks, particularly in adversarial environments. Further research into advanced aggregation techniques could enhance FTI-SLAM’s resilience against such threats.
The findings of this study, titled “FTI-SLAM: federated learning-enhanced thermal-inertial SLAM,” have been published in the journal Robot Learning, advancing the field of SLAM technology and its applications in sensitive and variable environments.
Source: Noah Wire Services