Motion Analysis and Object Recognition for Autonomous Navigation.
Abstract
The research in computer vision described in this final report is directed towards the achievement of autonomous vehicle navigation using passive visual sensing. For a modeled environment, we have implemented a navigation system incorporating reactive planning, and based on the identification of known landmarks in the 3D scene. Robust algorithms have been demonstrated for the recovery of pose--the position and orientation of the camera--from model matching between the image and known environment. For an unknown environment, a navigation system has been demonstrated in which image based homing is used to move between neighboring target locations. For a completely unknown environment, multi frame structure from motion algorithms have been developed which use image sequences for the reconstruction of the camera motion and environmental structure. In a partially modeled environment, the combination of pose recovery with triangulation over image sequences yields a robust, accurate algorithm for incremental acquisition of a 3D scene model. Lastly, a new framework for obstacle detection from motion has been developed and demonstrated experimentally. In the area of static image interpretation and object recognition, research has been done on perceptual organization, invariant features, 3D reconstruction, and the automatic learning of strategies for object recognition. We have developed a new approach to distinguishing figure from ground, a prerequisite for obstacle detection, based on perceptual grouping techniques.
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 01, 1992
- Accession Number
- ADA309654
Entities
People
- Allen R. Hanson
- Edward M. Riseman
Organizations
- University of Massachusetts Amherst