Adjustment of Point Pillars for Effective LiDAR-Based Object Detection and classification in Constrained Environments

Autonomous electric vehicles operating in confined environments face unique challenges, as the objects-within these environments often have specific characteristics-not adequately covered by general datasets. This condition is-particularly relevant for the autonomous electric vehicle being-developed by BRIN in Indonesia, which requires accurately-recognizing and classifying objects to avoid collisions and-navigate safely. To address these needs, our research proposes-the development of an object detection and classification-system utilizing the Velodyne VLP-16 LiDAR. Given the-tendency of this LiDAR to produce sparse point clouds, we-have adjusted the PointPillars method to better adapt to such-data, showcasing the adaptability of our system. Our findings-indicate that a backbone network model configuration based-on a residual network (BaseBEVRESBackbone) outperforms-the traditional PointPillars backbone configuration (BaseBEVBackbone). This superior performance is achieved-even with a more straightforward layer configuration and-smaller voxel size, demonstrating the effectiveness of our-approach in enhancing LiDAR-based object detection and-classification in constrained environments.