Deep Learning-based Real-Time Object Segmentation from 3D Point Cloud for Autonomous Driving

Abstract

Due to applicability of light detection and ranging (LiDAR) on various applications, semantic object segmentation from 3D point cloud generated by the LiDAR has attracted a lot of interests from researchers. Traditional approaches are based on handcrafted features and clustering algorithms to conduct 3D recognition from the point cloud. However, the approaches suffer from expensive computation and insufficient accuracy on predictions. To overcome these drawbacks, deep learning techniques are applied to semantically segment the point cloud by extracting features based on convolutional neural network (CNN) techniques. Despite the powerful ability of the CNN-based techniques on efficient features extraction, achieving high accuracy of segmentation on overall objects, especially small size objects, is still a challenging task. In this paper, we therefore propose an attention-based multi scale atrous convolutional neural network (AMSASeg) to solve the problems and provide accurate segmentation performance with real-time speed. The proposed network can obtain small to large information on overall objects (car, pedestrian and cyclist) and extract spatial to semantic features on small objects (pedestrian and cyclist). We evaluate the proposed method on KITTI dataset. Results shows that our network is effective to enhance segmentation performance on overall objects while satisfying real-time speed.