This paper proposes a method of obstacle detection based on 3D vision sensor, which is used to avoid false detection and missed detection of obstacle for robot navigation in quasi-static scenes. The algorithm firstly performed coordinate system conversion on the point cloud data obtained by the 3D vision sensor. Secondly, the depth image was generated with the height value as the scale factor, and then it was compared with the obstacle height parameter to generate the binarized image. Finally, we calculates the maximum connectivtity of the binarized image, and the size of field identifies the obstacle. The algorithm is used in the substation power inspection robot safety inspection project to verify the stability and effectiveness of the algorithm through long-term complex environment operation.