Masters Thesis: A vision-based motion estimator

For legged robots learning how to move efficiently requires an estimate of the motion that is being created, in particular velocity estimation is of importance. Because common sensors available for legged robots cannot provide good velocity estimates, the focus of this project has been on using visual sensors to measure this velocity. Because of their information-rich output, visual sensors offer an attractive, but challenging opportunity for velocity estimation. The goal of this graduation project was to develop an estimator for the motion of a legged robot, which outputs velocity using cameras as a sensor. After the literature study that has been performed in advance this project an additional goal has been defined. This subgoal is again inspired by nature, driven by the application of this project. If we were to compare the velocity estimation quality, what kind of animal with two eyes similar to cameras would excel? Would this be a prey-like animal with it’s eyes placed facing away from each other, or would it be a predator-like animal with it’s eyes facing the same direction? In the literature there is one particular method that showed good promise as a velocity estimator, since this method can also be used as a basis for an estimator for finding nature’s best eye setting. This is the filtering method for a single camera developed by Chiuso et.al. [1]. The method is a combination of two components, a system that represents the projection of points found in the camera image and a non-linear Kalman filter. The filtering method provides estimates of the pose, velocity and the locations of the observed points in the world. In this thesis an extension of the system introduced by Chiuso et.al. is proposed that represents the projection of points found in the camera images for an arbitrary number of cameras. For this thesis a setup with two cameras is defined, which can be used for both the predator and the prey setting. Similar to the filtering method for the single camera a non-linear Kalman filter is selected to be combined with the two-camera system, obtaining a velocity estimator for a two cameras setup. This new estimator has significant advantage over currently Master of Science Thesis David van der Lijn ii available estimators for finding nature’s best eye setting. With the new estimator, experiments can be performed for both the predator and the prey case, unlike the available estimators which require two different implementations for the two cases. In this project simulations have been used to approximate the real world application for velocity estimation using the new vision-based estimator. Experiments performed with these simulations have provided strong evidence in favor of the prey setting as nature’s best eye setting. Furthermore, experiments exploring the properties of the estimator have shown that the concept is working correctly. David van der Lijn Master of Science Thesis