Analysis and Noise Modeling of the Intel RealSense D435 for Mobile Robots

Abstract

Cameras that provide distance measurement along with RGB data have increasingly been appearing in the market as alternatives to the more expensive setup of LIDARs and webcams. While products such as the Kinect have existed in the past, its weight and form factor have been demanding constraints for mobile robots, specifically legged robots that are sensitive to payload. Recently Intel released a new lineup of Intel RealSense RGB-D cameras that have favorable characteristics for legged robots, specifically in terms of resolution, frames per second, form factor, weight, and price range. However, because these active stereo sensors are noisy for reasons such as non-overlapping image regions or lack of texture, it is beneficial to empirically model the noise. Systematic errors, specifically the distance inhomogeneity and depth bias, are observed to recognize and verify the limitations of the camera. We also analyze the non-systematic error by modeling both the axial and lateral noise as a function of distance and angle of incidence using a Gaussian distribution for its versatile applicability for mobile robots in mapping.