Abstract:
Robust perception is a critical prerequisite of autonomous robotics, enabling platforms such as unmanned ground vehicles (UGVs), aerial vehicles (UAVs), and surface vehicles (USVs) to interpret, navigate, and interact with complex environments. Traditional perception pipelines often depend on multi-sensor setups, including LiDAR, radar, or multi-camera rigs, which, while effective, introduce substantial costs, complexity, and calibration requirements. In contrast, single-sensor solutions based on monocular or omnidirectional cameras offer cost-efficiency, compactness, and ease of deployment.
This dissertation investigates how geometric reasoning and learning-based methods can be combined to achieve robust object detection, localization, and distance estimation, focusing on three complementary research directions: spatial memory, learning-based methods, and lightweight temporal fusion.
First, we introduce a memory-map framework for object detection and localization, using a single 360° camera, demonstrating that spatial confidence accumulation over time improves robustness to partial occlusions and inconsistent detections. Second, we extend this concept to UAV platforms, integrating onboard metadata such as GPS, IMU, and altitude into geo-referenced confidence maps. This metadata-driven approach enhances temporal consistency in aerial video object detection and provides interpretable localization across dynamic flight paths. Third, we propose learning-based formulations for object distance estimation. On USVs, we show that extending an object detector with a regression branch enables approximate distance prediction directly from monocular imagery, offering a cost-efficient alternative to conventional range sensors. Then we broaden this concept to omnidirectional 360° images, where the method generalizes across diverse datasets and challenging conditions, outperforming geometric and alternative learning baselines without explicit calibration. Finally, we present a lightweight multi-frame integration strategy for YOLO-based video object detection. By stacking consecutive frames and supervising only the final frame, the approach leverages temporal context with minimal computational overhead, significantly improving robustness against motion blur, occlusions, and appearance changes, while preserving near single-frame throughput.
Overall, this dissertation demonstrates that single-sensor vision systems, when enhanced with spatial memory, metadata-driven reasoning, learning-based distance regression, and lightweight temporal fusion, can deliver robust, efficient, and generalizable perception across ground, aerial, and maritime platforms.