To improve the depth estimation accuracy and efficiency of stereo vision, a stereo matching method based on binary disparity segmentation and 3D convolution is proposed. Firstly, features are extracted from the stereo image and fed into the segmentation module. Then, for each parallax plane, the 3D convolution module is run separately to detect whether the target is closer than the given distance, or to estimate the depth according to any rough order. Finally, 3D convolution layer is used to estimate the output of binary segmentation module, and the final disparity map is obtained after fine processing. Experimental results show that the 3-px error of the proposed method is 4.37% and the EPE error is 1.06 pixels in SceneFlow dataset. The small error approximation depth guided aggregation network (GA-Net) method on KITTI2015 dataset. And the proposed method has the highest efficiency in different depth quantization levels.