LSD-SLAM is a direct, feature-less monocular SLAM algorithm that enables the creation of large-scale, consistent maps of the environment. Unlike existing direct methods, LSD-SLAM allows for real-time, accurate pose estimation using direct image alignment and semi-dense depth maps. These depth maps are generated through filtering over many small-baseline stereo comparisons, enabling the system to handle large variations in scene scale. The algorithm operates on a CPU and is capable of running in real-time, even on a smartphone. The key innovations of LSD-SLAM include a novel direct tracking method that explicitly detects scale-drift and a probabilistic approach to incorporate noisy depth values into tracking. The system represents the global map as a pose graph of keyframes, with edges defined as 3D similarity transforms, allowing for the detection and correction of accumulated scale-drift. The algorithm uses a left-compositional formulation for pose estimation and incorporates uncertainty propagation to improve robustness. The algorithm consists of three main components: tracking new frames using direct image alignment, depth map estimation through filtering and regularization, and map optimization using pose graph optimization. The tracking component estimates the relative 3D pose of new images, while the depth map estimation component refines or replaces keyframes based on the current camera position. The map optimization component ensures consistency by detecting loop closures and scale-drift. LSD-SLAM has been evaluated on publicly available datasets and challenging outdoor trajectories, demonstrating its ability to accurately reconstruct 3D environments with large variations in scale and rotation. The system's scale-aware formulation allows for accurate estimation of both fine details and large-scale geometry, making it a versatile and robust solution for monocular SLAM. The algorithm's performance is validated through quantitative comparisons with other approaches, showing its effectiveness in real-time applications.LSD-SLAM is a direct, feature-less monocular SLAM algorithm that enables the creation of large-scale, consistent maps of the environment. Unlike existing direct methods, LSD-SLAM allows for real-time, accurate pose estimation using direct image alignment and semi-dense depth maps. These depth maps are generated through filtering over many small-baseline stereo comparisons, enabling the system to handle large variations in scene scale. The algorithm operates on a CPU and is capable of running in real-time, even on a smartphone. The key innovations of LSD-SLAM include a novel direct tracking method that explicitly detects scale-drift and a probabilistic approach to incorporate noisy depth values into tracking. The system represents the global map as a pose graph of keyframes, with edges defined as 3D similarity transforms, allowing for the detection and correction of accumulated scale-drift. The algorithm uses a left-compositional formulation for pose estimation and incorporates uncertainty propagation to improve robustness. The algorithm consists of three main components: tracking new frames using direct image alignment, depth map estimation through filtering and regularization, and map optimization using pose graph optimization. The tracking component estimates the relative 3D pose of new images, while the depth map estimation component refines or replaces keyframes based on the current camera position. The map optimization component ensures consistency by detecting loop closures and scale-drift. LSD-SLAM has been evaluated on publicly available datasets and challenging outdoor trajectories, demonstrating its ability to accurately reconstruct 3D environments with large variations in scale and rotation. The system's scale-aware formulation allows for accurate estimation of both fine details and large-scale geometry, making it a versatile and robust solution for monocular SLAM. The algorithm's performance is validated through quantitative comparisons with other approaches, showing its effectiveness in real-time applications.