Abstract
Sensors play a fundamental role in sensing the environment for autonomous vehicle perception, providing accurate and reliable data essential for understanding and navigating the surroundings. LiDAR sensors are widely used for their ability to generate detailed 3-D point cloud data of the surroundings. Bird’s-Eye View (BEV) detection utilizes these point cloud data to identify objects, such as cars and cyclists, from a top–down perspective. This LiDAR sensor-based perception approach is essential for understanding complex environments and ensuring safe navigation in real-time driving scenarios. This letter presents DSFNet, a compact LiDAR-only network for BEV perception. The model integrates an efficient pillar-based encoder with a proposed dual-scale fusion (DSF) backbone, designed to mitigate performance and complexity issues associated with LiDAR sensors. The backbone reduces parameter count by approximately 50% compared to standard architectures while maintaining competitive detection accuracy. By capturing both local detail and global context, DSFNet enhances feature representation for sparse and irregular LiDAR data. Evaluations on the official KITTI BEV benchmark demonstrate strong performance in car and cyclist detection, highlighting suitability for real-time sensor-driven applications.