Motion model binary switch for MonoSLAM

Abstract

Current Monocular Simultaneous Localization and Mapping (MonoSLAM) methodologies use constant velocity and smooth motion assumptions. If the motion consists of rapid accelerations, decelerations or stops, the position estimates become erroneous and unstable. Mobile robots require frequent stops due to mission dictated or safety reasons. With the objective of using MonoSLAM to localize a mobile robot, we determined the effectiveness of trajectory estimation for a typical robot moving with constant velocity and stopping to execute missions. Experiments were performed with a camera mounted on a 3-axis translational robot and several path profiles with brief stops were executed. The trajectory estimated with a MonoSLAM algorithm is compared with the known motion profile. As the stop causes significant error and drift in the position estimates, we modified the constant velocity motion model to incorporate a stop detection method. An optical flow based stop detection model was formulated and implemented in conjunction with MonoSLAM. Velocity update is modified when a stop or start is detected by optical flow. By adaptively switching between constant velocity and stop models, the trajectory estimate is seen to be more accurate and stable after an intermittent stop. Details of the adaptive switching method and the performance of the modified MonoSLAM are described in this paper.