Visual memory maps for mobile robots

In this paper, we introduce the concept of Visual Memory Maps (VMM). A VMM is an ordered accumulation of visual information, classified to avoid redundancy and to provide fast retrieval through the comparison of newly acquired images with stored memories. We use VMM to give some low level visual capabilities to mobile robots. To do this, a mobile robot is guided through a path while it takes snapshots. The image stream is then converted into a VMM. Two problems that surge are both the representation of the image stream and the comparison of new frames with the VMM. In this research, we use Principal Component Analysis to have a compact representation and to index the image stream, and introduce a registration scheme to search for correspondence directly in the space of the principal components. © 2000 IEEE.