Context From a stereovision acquisition, it is possible to build a D point cloud of the environment. For a moving cameras pair, we can use SLAM to merge the generated clouds of points, and create a dense D scene surrounding the sensors, without any other sensor than the two cameras. Figure: Point cloud generation from a stereo acquisition. To carry out this task on computationnaly limited devices, we offer a solution to efftiently reduce the number of processed frames without increasing the generated error. Adaptive frames selection Our algorithm focuses on strongly reducing the number of frames when the trajectory is mostly straight, and keeping a high frame rate during rotations. The gure below illustrates two outputs of the same SLAM algorithm with the same number of input frames with and without our adaptive frames selection. Figure: Adaptive frames selection. With our method (in black and red), the estimated trajectory ts perfectly the ground truth (in green). Whereas a naive frame selection (in blue) strongly moves away. Real-time Dense D Reconstruction Pipeline Our adaptive SLAM method lets us build a real-time dense 3D reconstruction process on a small device (such as a Raspberry Pi) without the need of a huge computationnal cost. Figure: Real-time dense 3D reconstruction pipelne. Figure: reconstructed point cloud correctly tting the GPS coordinates (in red).