Environment Classification for Indoor/Outdoor Robotic Mapping Jack Collier Defence R&D Canada – Suffield Autonomous Intelligent Systems Section PO Box 4000 Station Main Medicine Hat, AB, T1A 8K6, Canada jack.collier@drdc-rddc.gc.ca Dr. Alejandro Ramirez-Serrano Dept. of Mechanical & Manufacturing Engineering University of Calgary 2500 University Drive NW Calgary, AB, T2N 1N4, Canada aramirez@ucalgary.ca Abstract We present a novel perception system for mapping of indoor/outdoor environments with an Unmanned Ground Vehicle (UGV). The system uses image classification tech- niques to determine the operational environment of the UGV (indoor or outdoor). Based on the classification re- sults, the appropriate mapping system is then deployed. Image features are extracted from video imagery and used to train a classification function using supervised learning techniques. This classification function is then used to classify new imagery. A perception module observes the classification results and switches the UGV’s percep- tion system, according to current needs and available (reli- able) data as the UGV transitions from indoors to outdoors or vice versa. A terrain map that exploits GPS and Iner- tial Measurement Unit (IMU) data is used when operating outdoors, while a 2D laser based Simultaneous Localiza- tion and Mapping (SLAM) technique is used when oper- ating indoors. Globally consistent maps are generated by transforming the indoor map data into the global reference frame, a capability unique to this algorithm. 1 Introduction The ability of a UGV to interact effectively with its en- vironment depends on its ability to sense and interpret it, a process called perception. UGV perception is traditionally accomplished by combining range data from various sen- sors with localization data to create a geometric world rep- resentation that can be used to achieve a task. To increase the capabilities of today’s UGVs, robot perception capabil- ities must move beyond the purely geometric approach to gain a greater situational awareness. Increased perception capabilities will allow the UGV to better assess the envi- ronment and operate more effectively in a wider range of situations. In general, state of the art UGVs are designed to oper- ate within an assumed environment in which the parame- ters and constraints are well known. If these assumptions are valid, the UGV can operate effectively, but failure can occur when the assumptions are incorrect. It is desirable to have a perception system that can adapt to changes in its op- erational environment, thus extending its effectiveness and operational range. This paper details an adaptive perception system through the use of learning and vision algorithms. A system is im- plemented that allows the UGV to adapt its mapping system when transitioning between an indoor and outdoor environ- ment. The system uses image based features from a colour camera and supervised learning techniques to perform scene classification, adapting based on the classification results. In the case of the outdoor classification, a terrain mapping system is deployed. This system uses GPS and IMU mea- surements to provide an accurate pose estimate as well as stereovision range data to determine the geometric charac- teristics of the environment. When the UGV transitions to an indoor environment, where GPS data is no longer avail- able, the UGV employs a laser based SLAM system. In the SLAM process, point features, extracted from a 2D laser scan, are tracked as the robot moves and used to probabilis- tically estimate the UGV’s pose and the landmark locations estimates. This paper is organized as follows. Section 2 provides an overview of background technologies and algorithms rele- vant to this work. Section 3 details the proposed perception system. In Section 4, several experiments designed to vali- date the proposed system are detailed followed by a review of the proposed approach, a discussion of the obtained re- sults and future work in Section 5. 2 Previous Work Previously relevant research related to this work has fo- cused in two main areas, scene classification techniques that 2009 Canadian Conference on Computer and Robot Vision 978-0-7695-3651-4/09 $25.00 © 2009 IEEE DOI 10.1109/CRV.2009.6 276