Existing simultaneous localization and mapping (SLAM) systems based on neural radiation field technology perform well in static environments for positioning and rendering. However, in most dynamic scenes, neural implicit SLAM systems suffer from significant pose drift and struggle to meet real-time requirements. This paper proposes a real-time neural implicit-based visual SLAM system tailored for dynamic scenes. We differentiate foreground and background using semantic and depth information, then identify dynamic regions using optical flow. We update the motion status of feature points within dynamic regions using hypothesis testing and conditional probability to reduce the impact of highly dynamic objects on localization. We adopt static masks and use multi-resolution hashing encoding for real-time rendering to fill map voids caused by excluding dynamic regions. In pose optimization, we assign an optimization weight to each point to minimize the negative impact of dynamic points on global optimization. We conducted experiments on multiple publicly available dynamic datasets. The results indicate that our method outperforms other neural implicit SLAM methods in terms of tracking accuracy and mapping quality in dynamic environments.