Future Object Localization in Autonomous Driving Using Ego-Centric Images and Motions

In autonomous driving, future object localization (FOL) is actively used for trajectory prediction and a collision avoidance system. However, it is a challenging task to accurately determine the future locations of nearby pedestrians and vehicles during driving. In this paper, we propose a stochastic FOL using ego-centric images and motions (FOLe) that are generic information obtained from an autonomous agent. The proposed network consists of two staged sub-networks, including a future candidate network (FCN) and a future decision network (FDN) for localization. The FCN is used to generate several hypotheses to inform where an object will probably appear in an image according to its attribution. Our network directly produces the hypotheses in the future using only the ego-centric images and motions, which is trained with an end-to-end manner. The FDN predicts a multi-modal distribution based on the previous results of the FCN and determine the final location by maximizing the probability distribution. Experimental results demonstrate that the proposed model provides a superior performance to the state-of-the-art studies in nuScenes dataset.