Fast and Accurate LiDAR Semantic Segmentation

Fast and Accurate LiDAR Semantic Segmentation

Author Name : S. Brindha, T. P. Kamatchi, V. Gokul, R. Sathish Kumar, N. M. Sakthyram, B. Senoj

ABSTRACT

Perception in autonomous vehicles is commonly allotted through a set of various sensing modalities. Given the large quantity of overtly on the market labelled RGB knowledge and also the advent of high-quality deep learning algorithms for image-based recognition, high-level linguistics perception tasks area unit pre-dominantly solved victimization high-resolution cameras. As results of that, alternative device modalities probably helpful for this task area unit typically neglected. During this paper, we tend to push the state of the art in LiDAR-only linguistics segmentation forward so as to supply another freelance supply of linguistics info to the vehicle. Our approach will accurately perform full linguistics segmentation of measuring device purpose clouds at device frame rate. we tend to exploit vary pictures as Associate in Nursing intermediate illustration together with a Convolution Neural Network (CNN) exploiting the rotating measuring device model. To get correct results, we tend to propose a completely unique postprocessing rule that deals with issues arising from this intermediate illustration like discretization errors and foggy CNN outputs. We tend to enforced and completely evaluated our approach as well as many comparisons to the state of the art. Our experiments show that our approach outperforms progressive approaches, whereas still running on-line on one embedded GPU.