Indoor positioning systems using Ultra Wideband (UWB) achieve high positioning accuracy ( ${< }30$ cm). However, traditional localization approaches require many packet exchanges (e.g. two-way ranging) or challenging clock synchronization (e.g. time difference of arrival). To remedy this, we propose active fingerprinting using the channel impulse response (CIR) from a single UWB packet received at each UWB anchor. The proposed neural network anchor-subset selection method with Savitzky-Golay filter achieves a low mean absolute error ( $20.9 - 87.0$ cm), in contrast to signal strength based fingerprinting approaches that realize accuracies of $2 - 3$ m. Finally, with CIR interpolation the data collection overhead is reduced.