Abstract: Simultaneous Localization and Mapping (SLAM) utilizes visual sensor data to estimate camera poses for localization. In dynamic scenes, SLAM is often augmented with deep learning methods to improve localization accuracy. However, deep learning methods typically demand high computational resources, resulting in time delays during system operation that hinders real-time streaming processing. To address these challenges, this paper proposes a streaming perception-based visual SLAM method for dynamic scenes, which integrates a streaming evaluation metric that considers both localization accuracy and algorithmic time delay to reflect the system's streaming processing performance accurately. Based on this metric, the proposed method employs multi-thread parallelism and camera pose prediction to obtain continuous and stable camera pose outputs, enabling streaming perception-based visual localization. Experimental results demonstrate that the proposed method effectively improves the streaming performance of visual localization using deep learning methods in dynamic scenes.