When mobile intelligence agent performs the complex task of Simultaneous Localization And Mapping (SLAM), the interference of dynamic objects will weaken the correlation between feature points and the degradation of the system's localization accuracy. In this regard, the work aims to propose a visual SLAM algorithm based on YOLOv5 and geometric constraints for indoor dynamic scenes. First, based on YOLOv5s, the original CSPDarknet backbone network was replaced by a lightweight MobileNetV3 network, which could reduce parameters and speed up operation, and at the same time, it was combined with the ORB-SLAM2 system to obtain semantic information and eliminate a priori dynamic feature points while extracting ORB feature points. Then, the possible residual dynamic feature points were further culled by combining the optical flow method and epipolar geometric constraints. Finally, only static feature points were used for camera position estimation. Experimental results on the TUM data set showed that both ATE and RPE were reduced by more than 90% on average under high dynamic sequences compared with ORB-SLAM2, and the processing of one frame in the tracking thread took only 28.26 ms on average compared with the same type of systems of DS-SLAM and Dyna-SLAM, while guaranteeing localization accuracy and robustness. The algorithm can effectively reduce the interference caused by dynamic objects to the real-time SLAM process. It provides a possibility for more intelligent and automatic packaging process.