Initial Study of Neuromorphic Application for Vision-Based Navigation

Abstract

High precision landing capability increases accessibility to scientifically interesting areas and makes a safe landing. Vision-based navigation is one of the GN&C functions to achieve that. The navigation technique requires onboard and real-time image processing algorithms performing image recognition, feature extraction, matching, tracking, etc. However, despite the demand for fast image processing task, space-grade computers have been slower than commercial ones, lagging behind them by 1-2 orders of magnitude in performance. To address the issue, we study the applicability of a vision-based navigation algorithm for crater classification in architecture of neuromorphic processor chips which are expected to operate with ultra-low power consumption. The neuromorphic processor chip operates asynchronously and in parallel mimicking neuro-biological architecture like our brain. Therefore, the processor operates as a spiking neural network (SNN). The difference in mechanisms between timing-based spiking neurons and rate-based artificial neurons has a problem of how to make the SNNs with equivalent functions as artificial neural networks (ANNs). This study proposes a method to train the SNN by transferring weights of the ANN trained specially as synaptic efficacy of the SNN, and also evaluates the accuracy and power consumption of the SNN.