In this paper, we present a new method for detection of LSB matching steganography based on the local stationary characteristics of natural images. Under an additive noise steganography model, the noise variance of stego image is the sum of that of its corresponding cover image and that of stego-noise. As a result, the local variance histogram (LVH) will stretch rightwards after data embedding. The right-stretching effect is measured by the center of mass (COM) of the LVH. Second-order difference of the image is applied to enhance stego-noise and calibration technique is also applied to weaken the influence of image content. Experimental results show that the new steganalyzer outperforms prior art and provides relatively reliable results for embedding rates as low as 0.25 bits per pixel on uncompressed images.