In order to improve the robustness and real time performance of SURF based image matching algorithms, a new descriptor is proposed. We compute the new descriptor in a rectangle local region (the side set to 20s). Firstly, the local region is divided into 8 equal triangle subregion. Secondly, local region location grid is rotated to align its dominate orientation to a canonical direction. The keypoint dominate orientation and its orthogonalorientation is defined as the x and y directions of the descriptors local coordinate system.Thirdly, compute the Haar wavelet response in x and y directions within the keypoint local region. In order to reduce the boundary effect and outer noise, Haar wavelet response in the same Grid of different triangle is both assigned to each triangle in different weight, and then a gaussian weighting function is used. Compute the histogram of Haar wavelet response and absolute Haar wavelet response, so each triangle subregion constitutes a vector with 4 dimensions. Finally, a descriptor with 32 dimensions is constituted and the descriptor is normalized to achieve illumination invariance. The experimental results show that the performance of the new descriptor is even better than SURF descriptor.