The 1st Israeli Smart Transportation Research Center (ISTRC) Annual Conference
Szántó M, Vajta L.
A key component for the feasibility of a vision-based, distributed, low-latency mapping solution is the ability to reduce the amount of unnecessary data supplied by image acquisition sources. There are numerous techniques for the segmentation of useful information from images and masking the visual data flow accordingly. This presentation introduces a novel approach that uses Deep Learning (DL) for the segmentation and selection of objects that belong to the static and quasi-static regions of the sensed / depicted environment – and therefore useful for low-latency mapping tasks. The method presented here is an extension of our recently published solution – that was developed as part of our combined evaluation setup, which we call “CrowdMapping”. In this presentation, followed by a short outlook to other solutions published in the literature, we will introduce a comparison of our two methods using images and ground truth data provided by the KITTI benchmark suite.