KDepthNet: Mono-Camera Based Depth Estimation for Autonomous Driving 2022-01-0082

Object avoidance for autonomous driving is a vital factor in safe driving. When a vehicle travels from any random start places to any target positions in the milieu, an appropriate route must prevent static and moving obstacles. Having the accurate depth of each barrier in the scene can contribute to obstacle prevention. In recent years, precise depth estimation systems can be attributed to notable advances in Deep Neural Networks and hardware facilities/equipment. Several depth estimation methods for autonomous vehicles usually utilize lasers, structured light, and other reflections on the object surface to capture depth point clouds, complete surface modeling, and estimate scene depth maps. However, estimating precise depth maps is still challenging due to the computational complexity and time-consuming process issues. On the contrary, image-based depth estimation approaches have recently come to attention and can be applied for a broad range of applications. A vast majority of camera depth estimation methods intend to determine the depth map of the whole input image using binocular cameras or a 3D camera, which is time-consuming too. In this paper, a novel approach is proposed that predicts the depth of the head obstacle using only a 2D mono camera. The bounding boxes of barriers are extracted through a deep neural network at the first stage. Rather than those methods, which calculate the depth map of the entire image pixels, in this paper, the average depth of each bounding box is calculated and assigned as labels. Then labels and feature vectors (four values of the bounding box) are set as input data of the proposed method. This network maps feature vectors of the previous stage to the estimated depth values. The results suggest that the model can reasonably predict the depths of obstacles on the Kitti dataset.