Fusion of sparse direct method and feature-based approach in stereo point-line SLAM algorithm

The development of autonomous driving and augmented reality has led to increased demands for accuracy and speed in visual SLAM. Traditional feature-based visual SLAM performs real-time image analysis with high complexity and computation requirements, necessitating devices with exceptionally high arithmetic power. This poses challenges for deployment on low-power devices. Additionally, relying solely on photometric errors makes the system too sensitive to the environment. To tackle these issues, our algorithm integrates the feature-based method and the sparse direct method. We employ the sparse joint method for pose estimation in the tracking thread, considering both accuracy and speed. Incase of tracking failure, we extract feature point lines for tracking. Furthermore, to enhance system robustness in low-texture environments and fuse point-line features, we employ adaptive weighted extraction of point-line features, considering computational resources. Our experimental results, based on public datasets, demonstrate that this algorithm achieves faster operation speeds in simple environments and higher accuracy compared to other SLAM algorithms in complex environments and fast camera movement scenes.