A context-sensitive segmentation approach for outdoor terrain feature extraction

The thesis research is concerned with the general problem of vision based navigation. An important element of visual navigation is the localization process through which a navigating entity determines its current position on a map. Prior experiments with human subjects have identified a set of basic terrain features that humans exploit while performing the localization process. The research investigated computer vision techniques for automatically extracting a subset of these terrain features from both outdoor terrain imagery and from topographic maps. The ability to reliably extract these terrain features from imagery and maps of outdoor terrain represents an important step toward the development of systems that can navigate autonomously in an outdoor environment.; The research investigated the value of incorporating domain specific knowledge into the feature extraction phase of the scene understanding process. Contemporary feature extraction methods have fallen short of human performance levels when applied to maps and imagery of large scale terrain. Significant performance improvements are demonstrated by incorporating domain dependent contextual knowledge into the feature extraction process. The feature extraction results that were achieved using the developed domain dependent feature extraction methods are compared with those achieved using a conventional edge operator to demonstrate the benefit of using domain specific knowledge in the feature extraction process.; The topographic features extracted from terrain imagery provide one input to the localization process. Topographic features extracted from a terrain map provide the other input. This research investigated a novel technique based on a hydrologic flow simulation for extracting topographic features from digital terrain maps. The results are compared to those obtained through the application of an existing technique based on local slope characteristics of the terrain. The hydrologic flow based technique exploits global characteristics of the terrain and is shown to yield more accurate results than the local slope based technique. A 130 square mile area encompassing the Grand Teton National Park region was the subject of the investigation. A database of this region was established that contains both terrain imagery and a digital terrain elevation map of the region.