Robotic mapping and localization with real-time dense stereo on reconfigurable hardware

Abstract

A reconfigurable architecture for dense stereo is presented as an observation framework for a real-time implementation of the simultaneous localization and mapping problem in robotics. The reconfigurable sensor detects point features from stereo image pairs to use at the measurement update stage of the procedure. The main hardware blocks are a dense depth stereo accelerator, a left and right image corner detector, and a stage performing left-right consistency check. For the stereo-processor stage, we have implemented and tested a global-matching component based on a maximum-likelihood dynamic programming technique. The system includes a Nios II processor for data control and a USB 2.0 interface for host communication. Remote control is used to guide a vehicle equipped with a stereo head in an indoor environment. The FastSLAM Bayesian algorithm is applied in order to track and update observations and the robot path in real time. The system is assessed using real scene depth detection and public reference data sets. The paper also reports resource usage and a comparison of mapping and localization results with ground truth.