Low-cost high-speed sensor fusion with GRiSP and Hera

Details

Supervisors

Van Roy, Peter

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée

Abstract

Internet of Things applications using sensor fusion are increasingly numerous. Unfortunately, performing sensor fusion can require computational capabilities which can be a limiting factor for low-cost IoT devices. On top of that, some of these applications try to react in real-time to their environment making low latency a key element for them. These real-time applications need to perform sensor fusion at the edge in order to reduce latency and as fast as possible to react in time. Optimising the computations and the latency of these applications becomes their main challenge. Hera, a fault-tolerant low-cost sensor fusion framework, had been developed to provide users with a high-level ready-to-use sensor fusion tool. It has been developed and tested on GRiSP-Base boards, an Erlang based low-cost IoT device. However, Hera was slow and could not be used for real-time applications. In this master thesis, I have improved Hera's speed by porting it to GRiSP2, the new version of GRiSP boards, and by improving its matrix library. Initially, the matrix library of Hera was written in Erlang which has poor performances computations. Fortunately, Erlang offers the possibility of introducing C functions through the NIF mechanism. Using Numerl, a NIF matrix library, the matrix computations of the Kalman filter, the algorithm used for sensor fusion, were sped up by a factor of ten. The combined hardware and software improvements led to a new version of Hera, Hera v2.0, which is almost 25 times faster than originally.