2019/05/22 by Khaled Saleh, Saleh, Khaled, Ahmed Abobakr +9 · 1 citation
Computer Science · Engineering · Environmental Science · #Advanced Neural Network Applications #Autonomous Vehicle Technology and Safety #Remote Sensing and LiDAR Applications
paper · pdf · doi:10.48550/arxiv.1905.08955
Point cloud data from 3D LiDAR sensors are one of the most crucial sensor\nmodalities for versatile safety-critical applications such as self-driving\nvehicles. Since the annotations of point cloud data is an expensive and\ntime-consuming process, therefore recently the utilisation of simulated\nenvironments and 3D LiDAR sensors for this task started to get some popularity.\nWith simulated sensors and environments, the process for obtaining an annotated\nsynthetic point cloud data became much easier. However, the generated synthetic\npoint cloud data are still missing the artefacts usually exist in point cloud\ndata from real 3D LiDAR sensors. As a result, the performance of the trained\nmodels on this data for perception tasks when tested on real point cloud data\nis degraded due to the domain shift between simulated and real environments.\nThus, in this work, we are proposing a domain adaptation framework for bridging\nthis gap between synthetic and real point cloud data. Our proposed framework is\nbased on the deep cycle-consistent generative adversarial networks (CycleGAN)\narchitecture. We have evaluated the performance of our proposed framework on\nthe task of vehicle detection from a bird's eye view (BEV) point cloud images\ncoming from real 3D LiDAR sensors. The framework has shown competitive results\nwith an improvement of more than 7% in average precision score over other\nbaseline approaches when tested on real BEV point cloud images.\n