The RGBD semantic segmentation model based on the standard 2D convolution kernel mostly takes the depth map as a single channel.Due to the limitation of convolution kernel characteristics, the geometric structure information brought by the depth information cannot be fully exploited.To overcome this defect, this paper constructs depth-sensitive convolution kernels and a pooling layer to make rich mining of depth information,and uses depth-sensitive spatial pyramid pooling to extract multi-scale information, so as to realize the segmentation of objects of different scales.Results of experiment on NYU v2 and SUN RGB-D datasets show that this method effectively improves the overall semantic segmentation accuracy.