Detection and classification of anti-tank (AT) mines, buried in fine-grained dry sand, have been carried out using a hand held ground penetrating radar. Real AT mine bodies of types M47 and TMA4 are used with the explosive replaced by a surrogate. Interspaced metal rods and cylinder shaped bodies make false alarm objects. The sensor system consists of a horn-type antenna mounted on a swinging arm and fed by a stepped frequency radar emitter. An angular rate counter, ensuring records with reasonably accurate positions, monitors the arm deflection. A number of B-scans result in data sequences containing 75 measurements, each with 55 frequency samples in the 0.3-3 GHz band. The classification of targets relies on the shape of radar echoes returned as sampled waveforms. Given a certain control of the position of the sensor, a sequence of waveforms coming from different measurement points enhances the classification quality. Two classification methods have been used in parallel based on 2D matched filters. The first method uses 2D templates extracted from background subtracted B-scans known to contain prototype mines. These templates are matched to full B-scans using normalized correlation in the time domain. The second method uses time-frequency maps based on the pseudo Wigner-Ville transform. Here, the matching to data sequences is done by direct comparison of selected features. The results show that, combining both methods, perfect recognition is achieved for two different AT mines with very good discrimination against the non-mine objects.