Visual Recognition of Artifacts by Computer.
Abstract
This work describes a set of programs which implements a bottom-up artifact modeling and recognizing system. This system is implemented on a DEC KI-10 in BLISS-10, MACRO-10, and FORTRAN. Sequences of stereoscopic pairs of images are digitized, frame by frame condensed to vertex-string-surface graphs (V-S-S graphs), and re-encoded as half chunk graphs (H-C graphs). The single frame processing requires approximately 20 minutes per image. Single image derived H-C graphs are matched by stereo pairs for depth, and by time interval pairs for motion. By using depth cues, motion cues and intensity feature labels, individual object subgraphs are segmented. Individual object graphs are matched with and/or entered into an object graph library. The report describes two new edge detection algorithms, an edge-based region aggregation algorithm, a scan line oriented vertex-string encoding algorithm, a half chunk graph matching algorithm, and a histogram-based graph matching algorithm. The idea of the 'half chunk', an elemental curvature element which can be used to form scale and coordinate system invariant object graphs (or 'feature' centered object models) is introduced.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 01, 1979
- Accession Number
- ADA084998
Entities
People
- Charles Jeremiah Jacobus
Organizations
- University of Illinois Urbana–Champaign