A method in Curvelet transformation for human detection is proposed in this paper. First, the curvelet coefficient was obtained by conducting the second curvelet transform for images. Then several statistical characteristics were computed from each sub-band coefficient. Finally, according the possibilities of the statistical features chosen by AdaBoost weak classifiers, some reasonable statistical measures were chosen as the final edge feature vector. The experimental results on INRIA and MIT human database showed that CTD increased the detection accuracy comparing to HOG.