Autonomous Obstacle Avoidance And Navigation Of UAV Using Stereo Vision And3D Scanner

With the recent advancement in the UAVs/MAVs, it leads to more sophisticated and autonomouscontrol system enabling the system to complete the mission more precisely and reliably. Nowadaysthe major problem faced by the autonomous vehicle is obstacle avoidance and navigation in indoorcondition.This thesis presents the methodology for solving the problem of autonomous obstacle avoidanceand navigation of MAVs. In the proposed methodology we have explored different sensors andimplement multi sensor system for obstacle avoidance and MAV navigation. The initial system isdesigned with stereo vision, in which we did a lot of experiments on determination of camera model,being used for this purpose.The camera model which is used in the calculation of position and orientation of MAV is ofcritical nature. Simple solution from triangulation cannot be used for accurate depth calculation;hence we worked on this problem and find an empirical model, which produces much better results.Further finding and matching of feature points of stereo images has been done with SURF. Thematched feature points are then used for MAV position and orientation calculation.Real time avoidance of an obstacle with the help of stereo vision cannot be achieved simply withCPU. In this work we have shown two different methods to calculate the disparity map, one is DSPbased and other is GPU based. DSP based system is not so efficient as compared to GPU based, so thefinal implementation was done with GPU. Also the stereo matching algorithm based on variable sizesquare window has been investigated and an improved method based on variable shape and sizewindow is proposed. The new method improves the performance and eliminates most of the noisewhich cannot be achieved with the original method.A novel method is adapted for obstacle avoidance and navigation of MAV with Microsoft XboxGaming Sensor “Kinect”. Kinect is a laser scanner which produces3D map of the environment up to10meter deep. The method uses feature matching algorithm for relating two scans. The filteringcriterion has been designed to avoid any false matching of feature points and hence make the systemmore reliable and flexible. The navigation as well as obstacle avoidance system proposed in thisthesis have been tested in indoor condition and are found quite satisfactory. Stereo vision and3D laser scanner cannot detect transparent obstacles so to detect such obstacleultrasonic sensor has been used in parallel with the stereo vision and laser scanner. The ultrasonicsensor data is fused with the laser scanner data to achieve the goal of transparent obstacle detection.