Abstract:
This work presents the building of the pipeline for the autonomous vehicle (AV). Our task was to combine perception models and tracking algorithms to get real-time detection and tracking of objects in diverse road environments. We implemented two pipelines leveraging LiDAR and Camera inputs. For 3D LiDAR data, we utilize the VoxelNeXt detection model and the CenterPoint Tracker to process point clouds from three Velodyne VLP-16 sensors for multi-object detection and tracking. This pipeline predicts 3D bounding boxes and assigns unique IDs to detected objects. In parallel, we integrate YOLOP with ByteTrack to process image data captured by Intel RealSense cameras. YOLOP’s multi-task perception capability for object detection, lane detection, and drivable area segmentation is augmented by ByteTrack’s robust tracking algorithm, ensuring high accuracy even in complex dynamic environments
