Shape-based Machine Perception of Man-Made Objects on Underwater Sensor Data
暂无分享,去创建一个
This thesis focuses on the common case in underwater robotics where the objects that need to be perceived are man-made and known a-priori. The idea is that the information about the shape of the object can efficiently guide object localization, segmentation and classification algorithms to the correct result while mitigating the strong noise and occlusions present on underwater sensor data.
In this thesis different implementations of the general idea of using shape a-priori knowledge are investigated. First an efficient screening algorithm that finds potential target objects on synthetic aperture sonar images. In this thesis a fast integral image based template matching framework is described. New template types and feature types that take the shape of the objects of interest into account and that are tailored to the detection on synthetic aperture sonar images are introduced.
For the segmentation of noisy sensor data a superellipse fitting onto already extracted object contours is investigated. To this end, a novel linearisation of the fitting error equation was proposed. Experiments showed that the linearisation decreased the computation time significantly. A second approach towards segmentation of noisy sensor data was made by reformulating the active contours without edges framework to accommodate a superellipse shape constraint. With the new formulation it was also discovered that the implicitly assumed underlying Gaussian pixel intensity distribution can easily be substituted to a more fitting pixel distribution. This is especially beneficial when the imaging system is not an optical camera but for example a sonar or another challenging noisy imaging system.
The last method proposed in this thesis employs the a-priori known shapes to do a top-down classification verification via a new proposed method to approximately simulate the sonar image. The experiments show that the top-down approach is a valuable tool to identify false positives.