کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Enhancing Autonomous Vehicle Perception through Point Cloud Upsampling of LiDAR Data

نویسندگان :

Reza Mohammadi¹ Hanieh Taghizadeh Shivyari² Zahra Keshmir³ Mohammad Amin Mohseni pour⁴ Khosro Madanipour⁵

1- دانشگاه صنعتی امیرکبیر 2- دانشگاه صنعتی امیرکبیر 3- دانشگاه صنعتی امیر کبیر 4- دانشگاه صنعتی امیرکبیر 5- دانشگاه صنعتی امیرکبیر

کلمات کلیدی :

Point Cloud،LiDAR،Upsampling،Deep Learning،Autonomous Vehicles

چکیده :

Autonomous vehicles (AVs) rely on accurate perception to navigate safely in complex environments. LiDAR is a key sensing modality that provides detailed three-dimensional point cloud data. However, parts of this point cloud may be sparse, especially at long ranges, which limits critical tasks such as object detection. To address this inherent limitation, we propose a lightweight deep learning–based point cloud upsampling framework that enhances the density of sparse LiDAR data while preserving geometric consistency. The model employs a feature extractor with local feature extraction and two stacks of self-attention to capture spatial dependencies across points and improve contextual representation, A feature expansion with a NodeShuffle layer that increases the number of points[6], and a coordinate reconstruction layer that refines the output into accurate 3D coordinates. Our approach produces dense, high-quality point clouds that extend LiDAR’s effective sensing range, thereby improving the detection of small and distant objects. These improvements contribute to safer and more efficient autonomous driving.