Location Recognition Using a Very High Dimensional Feature Space

Abstract

This work is focused on creating an autonomous location recognition system that is capable of determining its location based on the percepts observed in the environment. This process involves segmenting the percepts in the region, segmenting the global region into local regions, developing models of the local regions based on the percepts present in that region, and recognizing both the percepts and regions. The models are based on the “dominant” percepts found in the global region, and are refined in order to define each local area. The feature space used to define the percepts is based on the hue, saturation, and value (HSV) color space quantized into a very high dimensional feature space (e.g. 10,000 dimensions). The global region is segmented into local regions using a relative perceptual difference measure between the current image and prior images. Once the local regions and global percepts have been found, the local models for each region are created and used for the location recognition process. Furthermore, a comparison of the current methods and prior methods of clustering the very high dimensional feature space are provided, as well as a comparison of the classification methods used based on this feature space. Finally, while the system moves through the environment, the percept blobs segmented are tracked and, based on their movement, defined. This involves recognizing reflections created by distant light sources, defining all other percepts with definitions ranging from actual percepts to aberrations of light, determining novel objects, and determining novel regions.