Simultaneous Localization and Mapping Systems for Dynamic Outdoor Environments

Abstract

Autonomous driving in dynamic outdoor environments demands robust and accurate perception systems capable of handling complex scenes with moving objects such as vehicles and pedestrians. This work presents advancements in Simultaneous Localization and Mapping (SLAM) systems tailored for such environments, focusing on the integration of dynamic object tracking, trajectory prediction, and real-time 3D scene reconstruction.

Building on visual-based SLAM methods, this research proposes a system that enhances localization and mapping accuracy by identifying and removing dynamic objects from the scene while optimizing camera motion and predicting object trajectories. Leveraging both 2D and 3D object detection methods, the system improves motion planning and collision avoidance, crucial for safe autonomous navigation. To overcome the inherent challenges of jitter and noise caused by rapid object movements, a lightweight trajectory prediction framework is introduced, enabling smoother and more reliable object tracking.
The study also explores the role of high-quality datasets in evaluating SLAM performance, emphasizing the need for dynamic, vehicle-based urban datasets that include moving pedestrians and cars. By using both real-world and simulated environments, the system is rigorously tested to ensure robustness across diverse conditions. Simulators such as CARLA and GAZEBO are employed to create comprehensive datasets that include ground truth for object trajectories, allowing for precise evaluation of visual odometry and mapping capabilities.
Furthermore, this work introduces a novel approach for improving SLAM in highly dynamic outdoor environments by separating the scene into static and dynamic components. Extending upon NeRF-based SLAM systems, the method utilizes octree structures for multi-resolution scene representation, achieving dense 3D reconstructions while filtering out dynamic foreground objects. The system demonstrates superior performance in reconstructing outdoor environments with fast-moving objects, providing more accurate and reliable localization and mapping for real-time applications.

Extensive evaluations on both simulated and real-world datasets validate the system’s efficiency, resilience, and accuracy, marking a significant step forward in the development of SLAM systems for dynamic outdoor scenarios, with promising applications in autonomous driving technologies.

Date of Award	7 May 2025
Original language	English
Awarding Institution	University of Bath
Supervisor	Wenbin Li (Supervisor)