IEEE Instrumentation and Measurement Technology Conference Anchorage, AK, USA, 21-23 May 2002 Abstract - This paper introduces an automatic approach for registration estimation between successive viewpoints of a laser range camera that takes advantage of the raw measurements and does not require any external device for pose estimation nor complex feature extraction or triangulation. Assuming only object rigidity and some overlap between the scan areas, the approach allows to estimate the six rotation and translation parameters that link 3-D scans gathered from different viewpoints. A compact modified Gaussian sphere representation is used to encode a simple planar patch approximation of the objects surface and to validate mapping between the measurements as the appropriate rotation and translation parameters are computed. This solution results in an important reduction of the computational workload and a sufficient accuracy for most robot navigation applications. The proposed approach is demonstrated in an experimental context using real range measurements collected from a series of viewpoints. I. INTRODUCTION Building virtual representations of 3D environments from range measurements requires that data are gathered from a large number of viewpoints. This requirement results from the complexity of objects to be modelled, from the limited field of view of sensors and from occlusions that occur between objects. Each dataset gathered from a given point of view is defined with respect to a local sensor-based reference frame. As a result, the sensor position and orientation at each viewpoint must be precisely estimated to ensure that the information obtained from every source is merged in a consistent way to build a 3D model. The problem of registration consists of determining the geometric relationship that exists between different views provided by the sensor. An imprecise registration between viewpoints prevents from computing reliable models for collision avoidance or fine interaction between a robot and its environment [11]. The sensor pose can be measured with external means such as magnetic position and orientation trackers, robotic arms or even CCD cameras providing images from which the sensor position and orientation can be extracted. The latter solution implies very complex image processing and pattern recognition algorithms that are time consuming and rarely fully reliable. The first two approaches appear to be more realistic. A magnetic position and orientation tracking device, such as the Fastrak system commercialized by Polhemus Inc. has been tested in our robotic workcell. Unfortunately, the magnetic fields used by the device to track the pose appear to be very sensitive to the environment. In an experimental setup containing quite a large number of metallic parts such as computer boxes, power supplies and robotic equipments, such a device does not succeed in providing the required pose information except in very limited circumstances and under constrained displacements. When a robotic arm is used to move the sensor from one viewpoint to another, the internal encoders of the robot also provide a good estimate of the sensor position and orientation. But our experiments revealed that there is still room for refinement on this information in order to enhance the quality of the virtual representation of the environment. Moreover, the sensor is then constrained to the robot physical workspace and cannot get an access to narrow areas of the environment. An interesting solution to estimate range sensor registration between successive viewpoints without any peripheral devices is to take advantage of the raw range data provided by the sensor. Assuming that there is an overlap between the areas of the scene that are measured from each viewpoint, it becomes possible to search for some matching characteristics in both sets of information and then compute the necessary registration information that would make the projections of those matching elements to superpose. In spite of the fact that the registration problem between range measurements has been studied for a while in computer vision, no extensive and definitive solution has been found yet. Many variations to the widely known iterative closest point (ICP) algorithm [1] have been proposed to match characteristic point sets [3, 10], curves, meshes [2, 4] or parametric surfaces [8]. Some of them use both range and intensity data, also provided by most range sensors, to improve their selection of control points that are to be matched [7, 12]. These algorithms generally provide good results but the search for characteristic curves or surfaces is very complex and time consuming. Moreover, research works on the topic of registration generally assume that full range images are directly available from the sensors. As a result, they search for matching characteristics between such full images and compute geometrical transformations from there. Such a framework does not correspond to the reality because the majority of range sensors currently available on the market or even prototypes found in laboratories do not provide such full images by themselves. They rather generate single points or scan lines of range measurements [6]. Those sensors that generate full images rely on an external mechanical device to translate the sensor or change its orientation [9]. This solution compares to the use of a robot to move the sensor and is sensitive in terms of registration errors. Scan-Based Registration of Range Measurements C. Chen, P. Payeur Vision, Imaging, Video and Audio Research Laboratory School of Information Technology and Engineering University of Ottawa Ottawa, Ontario, Canada [chenadiu,ppayeur]@site.uottawa.ca 0-7803-7218-2/02/$10.00 ©2002 IEEE