Long-distance three-dimensional object detection in rainy conditions from multi-sensory measurements using a new machine learning model

Abstract To address the challenges of feature extraction and low detection accuracy for long-range three-dimensional (3D) objects in rainy environments, this paper proposes an improved probabilistic and geometric depth (PGD) machine learning-based detection method using multi-sensory measurements. In this new method, the deformable convolutions replace the regular convolutions in the PGD network to enhance the feature layer adaptability to the target shapes. A receptive field enlarging network is also designed to improve the detection performance for distant objects. Furthermore, an optimized multi-level feature pyramid is developed to improve the detection accuracy across varying scales. Lastly, a decision-level fusion aggregates the PGD detection head and the branch network detection head using a deep fusion module to produce the 3D object information and recognize different targets in a long distance. The proposed method is validated using the KITTI, nuScenes public datasets, and real-world testing datasets, where the camera and millimeter-wave radar are used to collect these datasets. The analysis results indicate 4%, 5.4%, and 12% increase in detection accuracy over the original PGD algorithm across these three datasets, demonstrating superior feasibility and accuracy of the proposed method for long-range object detection.