An inverse perspective mapping technique based on a virtual-vertical plane model for obstacle detection using mono camera

This paper presents an improved inverse perspective mapping (IPM) technique based on a virtual-vertical plane model for obstacle detection. The conventional IPM based obstacle detection has a problem that it is difficult to determine whether the IPM based detected pixels are from real obstacle or caused by noise. Small errors in model, unknown camera motion, and colorful patterns of ground cause noises which are actually the ground region that is detected as an obstacle. It can be interpreted as false positives. Furthermore, there are few detections on homogeneous patterned obstacles such as an unicolor box. This work proposes a new simple method for distinguishing the noise and real obstacle in an image when IPM based detection method is used. A vertical plane model based detected pixel is checked after the IPM based detection. The results show that the real obstacles including a homogeneous patterned unicolor box are detected as obstacles, and A4-size black paper on a floor which is not an obstacle is not detected as an obstacle.