Probable Multi-hypothesis Blind Spot Estimation for Driving Risk Prediction

Abstract

We present a method for estimating free spaces and obstacles in blind spots occluded from a single view. Knowledge about blind spots helps autonomous vehicles make better decisions, such as avoiding a probable collision risk. It is essentially ill-posed to estimate whether unobservable areas are uniquely assigned as free or occupied spaces. Therefore, our framework is designed to be able to produce probable multi-hypothesis occupancy grid maps (OGM) from a single-frame input based on posterior distribution of blind spot environments. Compared to deterministic single result, each hypothesis OGM can show other probable environments explicitly even in uncertain areas. In order to handle this, we introduce a combination of generative adversarial networks (GANs) and Monte Carlo sampling. Our deep convolutional neural network (CNN) is trained to model an approximate posterior distribution with an adversarial loss and dropout layers. While activating dropout even at inference step, the network generates diverse multi-hypothesis OGMs sampled from the distribution by Monte Carlo sampling. We demonstrate that the proposed method estimates diverse occluded free spaces and obstacles in multi-hypothesis OGMs from either a two-dimensional (2D) range sensor measurement or a monocular camera image. Our method can also detect blind spots ahead of vehicle as driving risks in real outdoor dataset.