Geoffrey TaylorLindsay KleemanDepartment of Electrical and Computer Systems EngineeringMonash University, Clayton 3800Victoria, [email protected]@eng.monash.edu.auStereoscopic LightStripe Scanning:Interference Rejection,Error Minimizationand CalibrationAbstractThis paper addresses the problem of rejecting interference due tosecondary specular reflections, cross-talk and other mechanisms inan active light stripe scanner for robotic applications. Conventionalscanning methods control the environment to ensure the brightnessof the stripe exceeds that of all other features. However, this assump-tion is likely to be violated for a robot operating in an uncontrolledenvironment. Robust scanning methods already exist, but suffer fromproblems including assumed scene structure, acquisition delay, lackof error recovery, and incorrect modeling of measurement noise.We propose a robust technique that overcomes these problems, us-ing two cameras and knowledge of the light plane orientation todisambiguate the primary reflection from spurious measurements.Unlike other robust techniques, our validation and reconstructionalgorithms are optimal with respect to sensor noise. Furthermore,we propose a procedure to calibrate the system using measurementsof an arbitrary non-planar target, providing robust validation inde-pendently of ranging accuracy. Finally, our robust techniques allowthe sensor to operate in ambient indoor light, allowing color andrange to be implicitly registered. An experimental scanner demon-strates the effectiveness of the proposed techniques.Source code andsample data are provided in the multimedia extensions.KEY WORDS—depth perception, light stripe, active stereo,specular reflections, robustness, calibration, interference re-jection, optimal reconstruction, color registration1. IntroductionLight stripe ranging is an active, triangulation-based tech-niquefornon-contactsurfacemeasurementthathasbeenstud-ied for several decades (Shirai and Suwa 1971;Agin and Bin-ford 1973). By projecting a known feature onto the measuredThe International Journal of Robotics ResearchVol. 23, No. 12, December 2004, pp. 1141-1155,DOI: 10.1177/0278364904048831©2004 Sage Publicationssurface, active scanners provide a more robust solution to themeasurement problem than passive ranging techniques. Re-views of light stripe scanning and related range sensing meth-ods can be found in Hebert (2000), Bastuscheck (1989), andBesl (1988). Range sensing is an important component ofmany robotic applications, and light stripe ranging has beenapplied to a variety of robotic tasks including navigation (Ny-gards, Högström, and Wernersson 2000; Aldon and LeBris1994), obstacle detection (Haverinen and Röning 1998), ob-ject recognition for grasping (Alshawish andAllen 1995; Raoet al. 1989), and visual servoing (Khadraoui et al. 1996).The drawback of conventional single-camera light striperange sensors is that favorable lighting conditions and sur-face reflectance properties are required, so that the stripe canbe identified as the brightest feature in the captured image.However, the range sensor presented in this paper is intendedfor use on a humanoid robot operating in an uncontrolled do-mestic environment (Taylor and Kleeman 2001, 2002). Un-der these conditions, variousnoise mechanisms interfere withthe sensor to defeat conventional stripe detection techniques:smooth surfaces cause secondary reflections, edges and tex-tures may have a stripe-like appearance, and cross-talk canarise when multiple robots scan the same environment. Themotivation for this work was to develop a robust light stripesensor suitable for operation in these noisy conditions.A number of techniques for improving the robustness oflight stripe scanners have been proposed in other work, us-ing both stereo and single-camera configurations. Magee,Weniger, and Franke (1994) develop a scanner for industrialinspection using stereo measurements of a single stripe. Spu-rious reflections are eliminated by combining stereo fields viaa minimum intensity operation. This technique depends heav-ily on user intervention and a priori knowledge of the scannedtarget. Trucco and Fisher (1994) also use stereo cameras tomeasurea laser stripe, andtreat the system astwoindependentsingle-camera sensors. Robustness is achieved by imposing anumber of consistency checks to validate the range data, the11411142 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / December 2004most significant of which requires independent single-camerareconstructions to agree within a threshold distance. Anotherconstraint requires valid scanlines to contain only a singlestripe candidate, but a method for error recovery in the caseof multiple candidates is not proposed. Thus, secondary re-flections cause both the true and noisy measurements to berejected.Nakano, Watanabe, and Kanno (1988) develop a similarmethod to reject false data by requiring consensus betweenindependent scanners, but using two laser stripes and only asingle camera. In addition to robust ranging, this configura-tion provides direct measurement of the surface normal. Thedisadvantage of this approach is that each image only recov-ers a single range point at the intersection of the two stripes,resulting in a significant acquisition delay for the completeimage.Other robust scanning techniques have been proposed us-ing single-camera, single-stripe configurations. Nygards andWernersson (1994) identify specular reflections by movingthe scanner relative to the scene and analyzing the motion ofreconstructed range data. In Haverinen and Röning (1998),periodic intensity modulation distinguishes the stripe fromrandom noise. Both of these methods require data to be as-sociated between multiple images, which is prone to error.Furthermore,intensity modulation does notdisambiguatesec-ondary reflections, which vary in unison with the true stripe.Alternatively, Clark, Trucco, and Cheung (1995) use linearlypolarized light to reject secondary reflections from metallicsurfaces,based onthe observation that polarized light changesphasewith eachspecularreflection.However, thecomplicatedacquisition process requires multiple measurements throughdifferent polarizing filters.Unlike the above robust techniques, the method proposedin this paper uniformly rejects interference due to secondaryreflections, cross-talk, background features, and other noisemechanisms. The proposed algorithm solves the associationproblem