A Survey of Visual SLAM in Dynamic Environment: The Evolution From Geometric to Semantic Approaches

Abstract

Simultaneous localization and mapping (SLAM) is crucial for the progression of autonomous systems, including autonomous driving, augmented reality (AR), and robotics. Traditionally reliant on static environments, SLAM now confronts the complexities of the dynamic real world. Advancements in artificial intelligence (AI) and deep learning are propelling SLAM toward the enhanced management of these dynamics. This survey presents a comprehensive analysis of visual SLAM in dynamic settings, an area significantly advanced by semantic understanding and sensor fusion technologies. It begins with an examination of geometric SLAM methods, addressing their efficacy in static contexts and limitations amidst dynamic changes. Subsequently, it focuses on semantic-based SLAM techniques, emphasizing their capacity for nuanced environmental representation and dynamic object management. In addition, the survey explores cutting-edge multisensor fusion strategies that substantially improve SLAM’s robustness and precision in intricate environments. We offer a critical review of persistent challenges, including computational demands, sensor calibration, and the imperative for real-time processing. The survey concludes by identifying fertile areas for future research, highlighting the ongoing potential for SLAM technology innovation to adapt to the ever-changing environmental dynamics.