Noise characterization of depth sensors for surface inspections

Abstract

International audienceIn the context of environment reconstruction for inspection, it is important to handle sensor noise properly to avoid distorted representations. A short survey of available sensors is realize to help their selection based on the payload capability of a robot. We then propose uncertainty models based on empirical results for three models of laser rangefinders: Hokuyo URG-04LX, UTM-30LX and the Sick LMS-151. The methodology, used to characterize those sensors, targets more specifically different metallic materials which often give distorted images due to reflexion. We also evaluate the impact of sensor noise on surface normal vector reconstruction and conclude with observations about the impact of sunlight and reflexions

A Sensor Fusion Layer to Cope with Reduced Visibility in SLAM

“…The main conclusion obtained in their benchmarking experiments was that all the LRFs tested provide different levels of noisy and erroneous results with saturated outputs, which makes them almost unusable under these conditions. Similar conclusions were obtained by Tretyakov and Linder [21], and in a recent comparison presented in Pomerleau et al [22], in which the LRF Hokuyo URG-04LX, also used in the results reported in section 5.4.1, presents the highest values of disparity and error in depth measurements, among of all compared LRFs. As distinct to previously described works, we herein propose a multi-sensor approach based on a LRF and a sonar array which, despite being based on an affordable setup using only commercial off-the-shelf (COTS) sensors, can provide a robust solution when LRF measures are partially disturbed by the presence of particles that reduce visibility.…”

Analysis of digitized 3D mesh curvature histograms for reverse engineering

“…But noise characterization depends on the chosen 145 scanner. In fact, noise characterization for depth sensors [33] and for laser beams [34] are different. So, digitization noise estimation is difficult.…”

A Review of Point Cloud Registration Algorithms for Mobile Robotics

“…In this review, we limit ourselves to the second derivative given that a non-negligible noise level is expected from the sensor measurements and from the motion of the sensor. As shown in [Pomerleau et al, 2012a], even the first derivative needs a large supporting surface to overcome the sensor noise of typical laser rangefinders. For example, extracting the surface normal from a flat area of 10-cm radius already leads to an expected error of 1.6 • (0.03 rad) for the rangefinder Sick LMS-151.…”

“…When an error model is available, it is also possible to add noise information based on measurement distance, incidence angle, reflectivity, etc. Examples of noise models on distance reading are investigated for Sick LMS-151, Hokuyo URG and UTM in [Pomerleau et al, 2012a].…”