Detection, information fusion, and temporal processing for intelligence in recognition
- Carnegie Mellon Univ., Pittsburgh, PA (United States)
The use of intelligence in vision recognition uses many different techniques or tools. This presentation discusses several of these techniques for recognition. The recognition process is generally separated into several steps or stages when implemented in hardware, e.g. detection, segmentation and enhancement, and recognition. Several new distortion-invariant filters, biologically-inspired Gabor wavelet filter techniques, and morphological operations that have been found very useful for detection and clutter rejection are discussed. These are all shift-invariant operations that allow multiple object regions of interest in a scene to be located in parallel. We also discuss new algorithm fusion concepts by which the results from different detection algorithms are combined to reduce detection false alarms; these fusion methods utilize hierarchical processing and fuzzy logic concepts. We have found this to be most necessary, since no single detection algorithm is best for all cases. For the final recognition stage, we describe a new method of representing all distorted versions of different classes of objects and determining the object class and pose that most closely matches that of a given input. Besides being efficient in terms of storage and on-line computations required, it overcomes many of the problems that other classifiers have in terms of the required training set size, poor generalization with many hidden layer neurons, etc. It is also attractive in its ability to reject input regions as clutter (non-objects) and to learn new object descriptions. We also discuss its use in processing a temporal sequence of input images of the contents of each local region of interest. We note how this leads to robust results in which estimation efforts in individual frames can be overcome. This seems very practical, since in many scenarios a decision need not be made after only one frame of data, since subsequent frames of data enter immediately in sequence.
- OSTI ID:
- 466458
- Report Number(s):
- CONF-9610138-; TRN: 97:001309-0038
- Resource Relation:
- Conference: International multi-disciplinary conference on intelligent systems: a semiotic perspective, Gaithersburg, MD (United States), 21-23 Oct 1996; Other Information: PBD: 1996; Related Information: Is Part Of Intelligent systems: A semiotic perspective. Volume I: Theoretical semiotics; Albus, J.; Meystel, A.; Quintero, R.; PB: 303 p.
- Country of Publication:
- United States
- Language:
- English
Similar Records
Computer vision for locating buried objects
Computer vision for locating buried objects