3D Detection of Low-altitude Small Objects by Infrared and LiDAR Data Fusion

Low-altitude flight poses significant challenges for unmanned aerial vehicles (UAVs) in detecting and avoiding thin obstacles such as power lines and tree branches. To address this issue, an infrared image dataset comprising power lines and trees, which consists of 2912 images annotated with 14000 labels, is constructed, and an improved UAV-based object detection model based on YOLOv8 is proposed. The model incorporates a multi-scale perceptual feature fusion module (SE-EMA) to enhance the capability to detect multi-scale targets. Furthermore, an adaptive threshold focal loss (ATFL) function is introduced to optimize sample balance. Finally, based on the detection results, the radar point cloud and infrared images are fused to output the high-precision three-dimensional coordinates, distance information and semantic categories of target. The proposed model is verified based on our custom dataset. The proposed model achieves a 3.3 percentage point improvement in mAP@50-95 compared to the original YOLOv8-seg model, while maintaining a detection speed of 170 frames per second (FPS). When the proposed model is deployed on a test platform for 3D target distance measurement, the ranging error is less than 10 cm. These results demonstrate that the proposed model exhibits favorable adaptability and robustness in detecting and localizing the power lines and tree branches in low-altitude UAV environments. It effectively enhances the detection and localization accuracies of slender targets against complex backgrounds.