Stereo matching is one of the most active research areas in computer vision. As it is an ill problem, a perfect solution does not exist until now. Aiming at solving the problem of low accuracy and sensitivity to radiometric distortion caused by existing local matching algorithm, a new stereo matching algorithm using improved gradient cost and adaptive window is presented here. Besides of the magnitude information in traditional gradient cost, the phase information is introduced and the taw matching cost is transformed to eliminate outliers. An adaptive window for every pixel is constructed by utilizing the image structure and color intensity. An effective disparity refinement method based on local disparity histogram is employed, which gains disparity maps with high accuracy. The experimental results show that the proposed method has an average error of 6.1% in the Middlebury testing benchmark, while keeping strong robustness to radiometric distortion.