Simultaneous Mapping and Localization framework for drone
Introduction: This project is aimed at implementing a vision-based Simultaneous Localization and Mapping (SLAM) to genreate a 3D map of the environment using a GTSAM and iSAM packages from GTSAM matlab toolbox. The premise of the project is to find the survivors in a power plant that has been destroyed by a recent earthquake. Due to the hazardous materials, drones (quadcopters) will be used for this mission. There are two parts to the mission. First, building the map of the current condition of the plant using a light weight drone with nothing but a single camera and an IMU. Then use a second drone is capable of carrying other needed supplies for the rescue mission. The second drone will use the map from the first drone and go to the rescue target location with the supplies.
Algorithm: The algorithm for estimating 3D world co-ordinates of camera poses and AprilTags using GTSAM involves 3 steps:
-
Initialisation For GTSAM packages to optimize the world co- ordinates properly, the world co-ordinated are to be initialized properly. Knowing that bottom-left corner of tag 10 will be origin in world co- ordinates, the poses of the camera frames con- sisting of tag 10 can be computed with RPnP approach.
-
Building the Factor Map The factor graph is then built using GTSAM pack- ages for MATLAB, with the initial estimates, mea- surements in pixel co-ordinates and noise models.
-
Optimising the factor graph The factor graph, thus created, is optimized using Levenberg-Marquadt non-linear optimizer.
The iSAM part of the project too follows a similar approach. However, the factor graph are created and optimized in batches (usually 1 frame at a time), as opposed to all the frames at once.