At present, the vehicle obstacle detection system usually uses different devices or sensors to perceive and obtain the obstacle information. However, omni-directional obstacle detection is difficult to realize because these devices or sensors are usually easy to be affected by environmental lighting and the material properties of the obstacle surface. Furthermore, most sensors have limited information regarding distance, which limits their application to omni-directional obstacle detection. To solve this problem, this paper proposes a method using depth camera for omni-directional obstacle detection. A method applying region growth for depth image and a fast inpainting method for depth image are proposed to extract and repair the obstacle regions in the depth images obtained by installing depth cameras around the car body. An improved method applying iterative normalized cut is also proposed to cluster and segment fragmentary and irregular obstacle regions to generate the complete obstacle regions. Finally, the obstacle regions are overviewed using a three-dimensional visualization method to realize omni-directional obstacle viewing. The results of experiments conducted in an environment with different obstacles during the day and night demonstrate that, compared with other methods, our proposed approach can effectively promote the ability to detect complex obstacles, and largely improve the detection speed; furthermore, obstacle detection using our method will be unaffected by environmental lighting. Each of these advantages provided by our method can significantly promote the driving safety of unmanned or other types of vehicles.