Uncertainty-aware Monocular 3D Object Detection for Safe Driving Assistance

Safety is of utmost importance when utilizing mobile vehicles. The introduction of safe driver assistance and automated driving systems will not only solve this problem, but also reduce road accidents and make driving smoother. To achieve these goals, 3D object detection is an essential technology and various methods have been proposed. Among these, monocular 3D object detection is an important technique and it is very difficult to identify the dimensions, direction and position of an object from a monocular image. Previous methods have predicted the dimensions directly from monocular images, resulting in poor accuracy of the dimensions in the depth direction. However, this paper proposes an indirect method for estimating the dimensions in the depth direction by predicting the volume of 3D bounding box. Detection accuracy for distant and partially visible objects is improved by using a loss function that uses uncertainty to account for the variability in predictions that occurs during learning. The proposed method has received an excellent evaluation on KITTI dataset.