Abstract: Simultaneous localization and mapping (SLAM) plays an important role in the field of computer vision and robotics, and provides basic technical support for VR/AR application development. Traditional SLAM has difficulty and inefficiently handles scenarios with relatively simple and sparse features. In this paper, we propose an improved algorithm for reconstructing an arbitrary trajectory based on visual SLAM. The front-end thread employs a Hessian matrix for extracting and matching the feature point among neighboring frames. By applying affine transformation to regions of interest to rapidly identify feature points, the errors associated with point-cloud reconstruction and the camera trajectory are reduced. To improve the efficiency of camera trajectory optimization, the back-end thread optimizes a limited number of feature points. Local and global BA strategies are used to optimize the camera trajectory to reduce system errors, thereby improving the optimization efficiency. Moreover, 3D models can be introduced in real-time. The experimental results show that the improved visual SLAM algorithm achieves better real-time performance than the traditional SLAM.