This is an old revision of the document!
LSD-SLAM: Large-Scale Direct Monocular SLAM
Contact: Jakob Engel
In the last years - in particular with the advent of commodity depth cameras - dense methods for camera tracking and 3d-reconstruction have quickly gained popularity. Instead of abstracting the images to a set of feature observations (such as corners, blobs or line-segments), they direcly operate on the raw images. This allows to exploit all information in the image and achieve high-accuracy tracking and reconstruction results in real-time.
We aim at developing similar methods for ordinary monocular cameras, where no direct depth measurements are available. While the absence of a direct depth measuring means makes the problem more complex, it also has important benefits:
- The inherent scale-ambivalence allows to use image regions at any depth (the range of RGB-D cameras is very limited). Especially outdoors, where typically a large portion of the scene is far away, this is very beneficial.
- A monocular camera is significantly smaller, lighter and cheaper than stereo- or in RGB-D camera, and hence more suited as sensor for mobile or robotic applications.
Semi-Dense Visual Odometry
The proposed Semi-Dense visual odometry / SLAM method uses all image-regions which carry information, i.e., which have a non-negligible intensity gradient. In contrast to keypoint-based methods this includes regions with gradient in only one direction (edges). The statistical depth map representation and the omission of intensity-uniform image regions allows it to run in real-time on a CPU. In addition to highly accurate camera odometry, it provides accurate, semi-dense depth maps of the environment, which can be valuable e.g. for robot navigation.
Related publications
Export as PDF, XML, TEX or BIB