Embedded Event-based Visual Odometry

Abstract

This paper presents an event-based visual pose estimation algorithm, specifically designed and optimized for embedded robotic platforms. The visual data is provided by a neuromorphic vision sensor. The fully event-based proposed approach is based on Spiking Neural Networks and a modified Hough transform. The method is developed to detect a square visual feature. The multi-thread algorithm is implemented on a Raspberry Pi, the well-known single-board computer used on many embedded platforms, that is connected to a Dynamic Vision Sensor (DVS) through its USB interface. Validation is done on two different experimental platforms and highlights the ability of the odometry algorithm to determine the relative pose of a robot with respect to a square target, in the aim to be integrated in an event-based visual servoing in a future work.