Ground Vehicle Classification Using Multifrequency Multipolarization Resonance Radar

摘要

Experiments investigating the classification of ground vehicles using processed radar returns are described. The calibrated and scaled backscatter measurements of scale-model vehicles at several azimuth angles are used to establish a catalog representing the VHF resonance region radar returns of actual vehicle targets. The performance of both the nearest neighbor algorithm, (using frequency domain data), and a correlation algorithm (using time domain data), is investigated. The effects of wave polarization, azimuth angle, and other key parameters are examined. The consequences of introducing forced errors into the estimates of aspect angle are studied. A novel feature set employing the ratio of vertically and horizontally polarized radar returns is described, and its classification performance is examined. In general, classification is found to be very much dependent on the particular algorithm, azimuth angle and polarization of interest. The Nearest Neighbor algorithm, using Radar Cross Section Amplitudes as features, was found to perform quite well, yielding classification rates of about 90%, depending on target azimuthal angle and wave polarization. Increasing the number of classification frequencies improves performance, but only to a limit. Errors in aspect angle are found to significantly degrade classification performance. (Author)