Incremental object recognition using range sensors

This thesis presents a method for incrementally recognizing objects as they are scanned by range sensors mounted on a mobile platform, such as a construction, mining, or agricultural field robot. The method enhances the productivity of field robotic machines in these settings by allowing them to start planning and moving toward the object before scanning is complete, or execute other motion tasks without the need to stop and scan. The system consists of two components, an on-line method which accomplishes the recognition and an off-line method which generates a finite state machine and associated parameters which guide the process of incremental recognition. The on-line method handles range data from laser or radar range sensors and is robust to the noise and poor sensor data that can result from unmeasured sensor motion during scanning. The off-line method uses range data obtained by simulating a range sensor scanning the object model in a sequence of poses. The on-line component of the system is used by an automated machine to recognize and locate objects it must interact with in its work area. Since this method handles unmeasured sensor motion, costs of these automated systems can be reduced by eliminating the need for highly accurate positioning systems to compensate for motion during scanning.; For the purposes of conducting experiments and evaluating the results this thesis focuses on a specific instance of this problem, the recognition of objects used during excavation operations such as on-highway and off-highway trucks. Results are presented using range data from scanning laser and radar range sensors designed for the environment and tasks of large mobile equipment. Results are presented which show that with a single truck model the method can report incremental descriptions at a rate of 20 Hz. This method has been used in demonstrations in which a hydraulic excavator equipped with range sensors and on-board computing autonomously loads multiple trucks.