Target Pose Estimation from Radar Data Using Adaptive Networks
Abstract
This research investigates and extends recent work by J.C. Principe at the University of Florida in target pose estimation using adaptive networks. First, Principe's technique is successfully extended to estimate both azimuth and elevation using SAR images. A network trained and tested using MSTAR data yields mean errors of less than six degrees in azimuth and five degrees in elevation. Second, the technique is applied to high-range resolution radar (HRR) signatures. Ground target (azimuth only) testing yields mean errors of less than 11 degrees for most classes. Air target testing for networks trained and tested on the same aircraft class yields mean errors under five degrees in azimuth and six degrees in elevation. However, cross-class estimation yields poor results. HRR signatures are analyzed to identify error sources and recommend ways to improve accuracy. Finally, Principe's novel mutual information training method is compared against traditional mean-squared-error training. Results show both methods are generally equivalent, but mutual information experiences convergence problems for some complex training sets. In general, adaptive network techniques demonstrate significant potential for improving the state of the art in target pose estimation. Both the estimation of elevation in SAR and the application to HRR are new and noteworthy successes.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 1999
- Accession Number
- ADA361658
Entities
People
- Andrew W. Learn
Organizations
- Air Force Institute of Technology