A U-Net Based Drivable Region Detection Algorithm for RGB-D Maps

To eliminate the influence of poor light conditions during drivable region detection and improve network based road edge contour detection accuracy, RGB-D-U-CRF, aU-Net based drivable region detection algorithm was proposed here in combination with RGB and maps. The proposed algorithm is featured with a dual branch late fusion structure. At different branches, image information of diverse forms was extracted and then subjected to late fusion; and then conditional random fields could be constructed according to RGB-D maps to optimize the boundary of the detected region. It is experimentally demonstrated that RGB-D-U-CRF is capable of performing fine detection of drivable regions in multiple poor light conditions including exposure, reflection and shades in a Cityscapes data set. In comparison with U-Net that merely adopts RGB images, pixel accuracy and intersection over union (IoU) are respectively raised by 5.88% and 6.16%. Besides, the detection time of the proposed algorithm is figured out to be 187ms in this study, signifying a good detection effect.