Streaming Perception-Based Visual Localization Method in Dynamic Scenes

In dynamic scenes, simultaneous localization and mapping (SLAM) is usually combined with deep learning methods to improve the localization accuracy of the system. In order to solve the problem of time delay generated during the operation of deep learning methods, which makes it difficult for the system to meet the requirements of streaming processing, a streaming perception-based visual SLAM method in dynamic scenes is proposed. Firstly, to address the limitation of traditional evaluation metrics that only consider localization accuracy, a streaming evaluation metric is proposed, which considers both localization accuracy and time delay. This metric can accurately reflect the streaming processing performance of the system. Secondly, to address the incapability of traditional visual SLAM methods to enable streaming processing, a streaming perception-based visual SLAM method is proposed, which combines multi-thread parallelism and camera pose prediction to obtain continuous and stable camera pose outputs. The evaluation results based on the BONN datasets and streaming evaluation method show that, compared with DynaSLAM, the absolute trajectory error (APE), relative translation error (RPE_trans), and relative rotation error (RPE_angle) of the proposed method have decreased by 80.438%, 56.180% and 54.676% respectively. In real-world scenes, the proposed method can obtain camera trajectories that are consistent with reality.