Time domain correlation technique is a widely used method for blood flow velocity measurement. The time shift between a pair of windowed ultrasonic echoes is estimated by searching the temporal position of the maximum of the interpolated normalized correlation function. Between two consecutive echoes, the acoustical footprint of a group of scatterers, which are transported with the flow, moves and is deformed. This implies a decreasing of the amplitude of the normalized correlation coefficient. In the case of microcirculation (low flow rate, low SNR), the amplitude of the correlation peak can be used to detect the presence of blood flow and to discriminate false and true detections (reliability index). We have numerically evaluated the statistical performances of the cross-correlation algorithm used as a correlation peak amplitude estimator in severe conditions (short correlation window length, low SNR). These theoretical results have been compared with in vitro experimentation on a 100-/spl mu/m-diameter microcirculatory phantom and with in vivo experimentation on a 180-/spl mu/m-diameter vessel of a human leg carrying erysipelas.