2020/04/02 by Prannay Kaul, Daniele De Martini, Kaul, Prannay +5 · 3 citations
Engineering · #Robotics and Sensor-Based Localization #Indoor and Outdoor Localization Technologies #Geophysical Methods and Applications
paper · pdf · doi:10.48550/arxiv.2004.03451
This paper presents an efficient annotation procedure and an application\nthereof to end-to-end, rich semantic segmentation of the sensed environment\nusing FMCW scanning radar. We advocate radar over the traditional sensors used\nfor this task as it operates at longer ranges and is substantially more robust\nto adverse weather and illumination conditions. We avoid laborious manual\nlabelling by exploiting the largest radar-focused urban autonomy dataset\ncollected to date, correlating radar scans with RGB cameras and LiDAR sensors,\nfor which semantic segmentation is an already consolidated procedure. The\ntraining procedure leverages a state-of-the-art natural image segmentation\nsystem which is publicly available and as such, in contrast to previous\napproaches, allows for the production of copious labels for the radar stream by\nincorporating four camera and two LiDAR streams. Additionally, the losses are\ncomputed taking into account labels to the radar sensor horizon by accumulating\nLiDAR returns along a pose-chain ahead and behind of the current vehicle\nposition. Finally, we present the network with multi-channel radar scan inputs\nin order to deal with ephemeral and dynamic scene objects.\n