Using Optical Flow to provide robustness for animage sensor based multiangulation positioning system

Sammanfattning

The use of image sensors in positioning systems has increased significantly the last years, spanning from complex SLAM-systems to simpler triangulation based methods. In the latter case, occlusion of anchor points may render the triangulation unsolvable, and any practical implementation must address this problem. This thesis aims to investigate if optical flow can be used to improve robustness against occluded anchor points in an image sensor based multiangulation positioning algorithm. Two fallback methods are implemented and evaluated with respect to their behavior in an indoor environment while being subjected to different types of occlusions to simulate real-world situations. The two considered methods are optical flow and a three-dimensional extension of a benchmark model presented by L. Martón et al. [26]. Evaluation is done on a small prototype UAV platform, while processing both simulated and collected datasets. The results show that the benefits of the optical flow-based algorithm are that it can handle complete and frequent occlusions in the cases where the occlusion time is short. Performance decreases with longer occlusions due to the drift of the estimation. Polynomial regression show a higher initial error but lower drift which makes it more suitable for longer, partial occlusions. In the performed experiments, the cutoff point where the error of the Visual Odometry estimation grows larger than that of Polynomial Regression is at about 15 frames at which point the horizontal positioning error is about 2.5% of the vertical distance to the ground.