Urban object detection using sensor fusion for autonomous navigation

Abstract

Urban object detection remains an important challenge for autonomous naviga tion, particularly in accurately identifying and locating thin and small structures such as poles. These structures are often difficult to detect due to their size, shape, and integration into complex urban scenes. In response to these challenges, this thesis proposes a dual-stage approach to urban object detection by integrating 2D deep learning-based object detection into 3D spatial representation. The methodology consists of two main stages: (1) 2D object detection, which involves fine-tuning a pretrained model to identify pole-like objects in urban environments, and (2) 3D object detection, where the 2D detections are projected into 3D space and further processed through point cloud clustering and outlier rejection to generate accurate 3D bounding boxes. The approach shows how combining segmentation-derived labels with geometric modeling can serve as an alternative to fully supervised 3D object inference. The model demonstrates promising performance on a custom dataset, benefiting from the SAHI framework, which enhances the detection of smaller or more challenging to-see objects. These results indicate that integrating 2D and 3D detection methods can provide valuable spatial information, even when detailed 3D annotations are limited or unavailable, supporting practical urban navigation applications. This thesis presents a foundational pipeline for pole detection in urban naviga tion, laying the groundwork for more adaptable and robust autonomous systems in complex real-world environments. Further development, including dataset growth and model refinement, can enhance detection accuracy and robustness, thereby ad vancing the effectiveness of autonomous systems in urban environments.

9