Pipeline for Automation of LiDAR Data Annotation

Light Detection and Ranging (LiDAR) sensor plays a vital role in the area of autonomous vehicles. LiDAR not only provides depth information of the object but also provides three-dimensional geometry. The task of object detection can be directly performed on LiDAR data, thus complementing the camera. However, to train the deep learning models for LiDAR object detection, the acquired LiDAR data must be annotated. LiDAR data annotation task includes creating a cuboid-shaped bounding box around objects like cars, trucks, pedestrians, etc. In this paper, we present a generic automation pipeline that eases LiDAR data annotation's tedious task. The pipeline was implemented on MATLAB using the LiDAR Labeler application for Car, Truck, and Pedestrian object classes. The proposed pipeline provides flexibility to add more automation labels as per the requirement. Another challenge in LiDAR data annotation is the non-adaptability of annotation tools for data acquired from other LiDAR sensors. A preprocessing step has been applied to raw data to address this problem. The pipeline was tested by annotating 50 point clouds from Pandaset data, and more than 80% of Cars and Trucks were found to be correctly annotated.