Received Sylff fellowship in 2024
Academic supervisor: Hideyuki Kawashima
Current affiliation: Keio University Graduate School of Media and Governance
In the near future, a society where a variety of mobility systems, including cars, delivery drones, and robots, operate throughout cities will become a reality. Leveraging the vast amounts of data collected from these mobility systems will enable the development of technologies such as large-scale autonomous driving, collision avoidance, and traffic congestion mitigation. To support these next-generation mobility technologies, I have conducted two research projects focusing on data systems tailored for mobility applications.
The first project involves the development of a system that efficiently manages the location information of a large number of mobility systems. Currently, autonomous vehicles and robots utilize a technology called Transform Library (TF) to manage location data. However, TF faces challenges in efficiently handling high-access requests. I applied database technologies optimized for high-access environments to TF to address this issue, resolving this limitation. Currently, I am expanding the system’s capabilities by incorporating distributed processing and redundancy mechanisms to enhance its reliability.
The second project focuses on a mechanism that uses images and point cloud data for object recognition through machine learning and registers the recognized data into TF. Machine learning is essential for mobility systems to recognize locations accurately. Although recent machine learning models have achieved dramatic performance improvements, they pose challenges such as increased memory usage, computational overhead, and significant power consumption. To tackle these issues, I have been researching quantization techniques that reduce the internal data representation of models from the conventional 16-bit to as low as 1.58-bit, addressing these challenges effectively.
Notably, these two research projects are not independent but closely interrelated. The location management system developed in the first project serves as the foundation upon which the object recognition mechanism in the second project operates, highlighting the importance of co-designing both systems.
