IEEE SENSORS JOURNAL 1 Towards Refocused Optical Mouse Sensors for Outdoor Optical Flow Odometry Robert Ross, Member, IEEE, John Devlin, Member, IEEE, and Song Wang Abstract—This paper investigates the use of refocused optical mouse sensors for odometry in the field of outdoor robotic navigation. Optical mouse sensors like the ADNS-2610 are small, inexpensive, non-contact devices, which integrate a CMOS camera and DSP hardware to provide two-dimensional optical displacement measurements. Current research indicates that vertical height variance contributes as a dominant cause of systematic error to horizontal displacement measurements, which raises significant problems for irregular environments encoun- tered in outdoor robotic navigation. In this paper we propose two approaches to mitigate this systematic error induced by height variance. The efficacy and robustness of the proposed approaches are tested by experimentation on an asphalt concrete road surface and by simulation. Index Terms—Optical Mouse, Optical Flow Odometry, Out- door Navigation, Dead Reckoning, Robot Navigation I. I NTRODUCTION O NE of the key questions which needs to be asked by mobile robots is “Where am I?”. Accurate localisa- tion information allows mobile robots to accurately perform mapping, obstacle avoidance and path planning functions [9]. Several different techniques have been developed to allow a robot to ascertain its location. Typically these can be classified as absolute (reference based) or relative (dead-reckoning) positioning techniques [4]. Absolute positioning uses beacons (like satellites for Global Positioning System (GPS)) or feature extraction to infer the current location based on specific cues (signals or specific objects) which help to localise the robot. In contrast, relative positioning, including inertial and wheel odometry, infers the current position based on the accumulation of movements from an initial position. Hence, the current position at any time is specified relative to the starting position. One widely used from of relative positioning, particularly on aircraft but also on mobile robots, is an Inertial Navigation Systems (INS) [4]. Commercial Micro-Electro-Mechanical Systems (MEMS) INS systems, where miniature cantilevers and springs are used to create minute accelerometers and gyroscopes, are showing increasing low navigation error in the order of a few percent [7]. These errors accumulate from the small errors which are magnified in the double integration required to convert from acceleration data to positional data [4]. R. Ross, J. Devlin and S.Wang are with the Department of Electrical and Computer Engineering, La Trobe University, Melbourne VIC, 3086 Australia e-mail: R.Ross@Latrobe.edu.au Manuscript received May 2, 2011 In contrast to the typical size of mobile robots, absolute positioning techniques (like traditional GPS) feature low res- olution (in the order of several metres for commercial GPS (or more when obstacles cause multipath reflections) [24] [21]) or in the order of centimetre resolution for Assisted- GPS with local active beacons [18] or fiduciary markers, and thus have limited usefulness for fine navigation over short distances in an unknown environment [5]. In addition GPS can be jammed or is rendered unavailable when signals from the satellites are blocked. To provide higher resolution without jamming or unavailibility constraints, relative positioning is traditionally used in the form of wheel odometry, where the displacement of the robot is inferred by measuring the amount that each wheel has rotated [4]. Due to the kinematic nature of wheels, particularly with their contact and slippage on different surfaces, an error in both orientation and position will continue to accumulate as the robot navigates an environment and performs path integration to infer its position [5], [13], [22]. Optical mouse sensors have been suggested as an alterna- tive method of performing odometry by using the sensors two-dimensional displacement output; see e.g. [12], [15], [17]. Optical mouse sensors integrate a small Complementary metal-oxide semiconductor (CMOS) camera (of the order of 18x18 pixels) with some Digital Signal Processing (DSP) and proprietary firmware algorithms to infer the displacement in both X and Y directions, based on the optical flow of features identified in consecutive image frames [1]. The concept of optical mouse sensors for odometry has clear advantages over using wheel encoders, such as the independence with respect to kinematic forces (including wheel slippage) and movement being resolved in more than one axis, which is of particular interest for omni-directional platforms [6]. Several sources of error when using optical mouse sensors are identified, namely the type of surface, height variance, lighting conditions and the angular displacement [14], [17]. Due to these sources of error, particularly height variance, most existing research into odometry with optical mouse sen- sors is being performed in well-controlled indoor environments on smooth surfaces, an implementation which is impractical for outdoor robotic navigation. In this paper we aim to develop robust dead-reckoning approaches for outdoor mobile robots with the optical mouse sensor assembly decoupled from the asphalt concrete test surface. The remainder of this article is organised as follows. Section II discusses current research whilst outlining the difficulties in using existing techniques for odometry in outdoor mobile robots. In Section III, we propose two approaches which can