Egocentric Vision-based Future Vehicle Localization for Intelligent Driving Assistance Systems

Predicting the future location of vehicles is essential for safety-critical\napplications such as advanced driver assistance systems (ADAS) and autonomous\ndriving. This paper introduces a novel approach to simultaneously predict both\nthe location and scale of target vehicles in the first-person (egocentric) view\nof an ego-vehicle. We present a multi-stream recurrent neural network (RNN)\nencoder-decoder model that separately captures both object location and scale\nand pixel-level observations for future vehicle localization. We show that\nincorporating dense optical flow improves prediction results significantly\nsince it captures information about motion as well as appearance change. We\nalso find that explicitly modeling future motion of the ego-vehicle improves\nthe prediction accuracy, which could be especially beneficial in intelligent\nand automated vehicles that have motion planning capability. To evaluate the\nperformance of our approach, we present a new dataset of first-person videos\ncollected from a variety of scenarios at road intersections, which are\nparticularly challenging moments for prediction because vehicle trajectories\nare diverse and dynamic.\n