Simultaneous Localization and Mapping (SLAM) in Dynamic Environments: State-of-the-Art and Simulation Benchmarking

Abstract

Traditional SLAM algorithms assume static environments, yet real-world unmanned systems must navigate spaces populated with moving objects—pedestrians, vehicles, and other robots—requiring dynamic SLAM approaches that distinguish static structure from transient elements. This paper provides a comprehensive state-of-the-art review of dynamic SLAM methods and presents extensive simulation benchmarking across diverse scenarios. We systematically categorize dynamic SLAM approaches into geometric methods that detect motion through multi-view geometry, semantic methods leveraging object recognition to identify potentially dynamic entities, and learning-based methods that implicitly model scene dynamics through neural networks. The review examines detection strategies for dynamic objects including motion segmentation via optical flow, background subtraction techniques, and deep learning-based instance segmentation, analyzing their computational complexity and accuracy trade-offs. We evaluate tracking and mapping strategies that handle dynamic elements: filtering approaches that reject dynamic features, dual-layer maps separating static and dynamic components, and probabilistic frameworks that model object motion patterns. Particular emphasis is placed on semantic SLAM systems integrating object detection networks (YOLO, Mask R-CNN) with traditional SLAM pipelines, examining how semantic understanding improves robustness in crowded environments. The paper presents a comprehensive simulation benchmarking framework implementing twelve prominent dynamic SLAM algorithms across standardized datasets and custom scenarios with controlled dynamic content. Benchmark environments span indoor navigation with pedestrian traffic, urban driving with vehicular motion, and warehouse operations with mobile robots, systematically varying the proportion of dynamic content (10-70% of visible scene). Performance evaluation encompasses localization accuracy, map consistency, computational efficiency, and robustness to different motion patterns (linear, erratic, occluding). Simulation results reveal that semantic approaches achieve superior accuracy in highly dynamic scenes but require significant computational resources, while geometric methods offer better efficiency with degraded performance as dynamic content increases. We investigate failure modes including dynamic object misclassification, map corruption from undetected motion, and tracking loss during rapid scene changes. The benchmarking examines sensor modality impacts, comparing camera-based, LiDAR-based, and multi-modal dynamic SLAM across varying lighting and weather conditions.