Visual Odometry Based Navigation Methods For Unmanned Aerial Vehicles

Autonomous navigation is the basis of unmanned aerial vehicles(UAVs)to accomplish various tasks.In order to achieve reliable autonomous navigation in large and unknown environments,research on navigation methods without GPS and based on vision is of great theoretical and practical significance.In this thesis,based on a visual odometry algorithm,visual navigation is achieved on a quadrotor platform.Considering the fact that the visual odometry algorithm might fail,two recovery approaches that work in different flight scenarios are proposed.In the end,several real flight experiments are conducted to verify the proposed methods.The main work of the thesis includes:1.Based on a visual odometry algorithm,a visual navigation system is implemented on a self-developed quadrotor platform.Implementation details include: data communication between visual navigation program and flight control program,scale calibration based on a statistical iterative method,horizontal plane alignment based on visual estimation,a complementary filter of vehicle position,etc.Outdoor flight experimental results demonstrate that the system can achieve autonomous navigation for UAVs in unknown environments.2.In terms of the problem that the visual odometry algorithm fails in scenes where image features are sufficient,a failure detection and reinitialization strategy is proposed.During the failure,the pose of the UAV is estimated based on two-view geometry.By using the map points in the two maps before and after reinitialization,scale correction that does not rely on other sensors is achieved,thus ensuring the consistence of the scale.Outdoor flight experimental results of the quadrotor show that the proposed method can ensure successful reinitialization after the visual odometry algorithm failure.3.In terms of the problem that the visual odometry algorithm fails in scenes where image features are insufficient,a reinitialization strategy based on optical flow is proposed.Besides,a motion model based on optical flow is established,and the visual speed is estimated.Then,the visual scale and actual speed of the vehicle is estimated based on an extended Kalman filter which fuses acceleration measurements,thus vehicle navigation is achieved during the visual odometry algorithm failure.Outdoor flight experimental results of the quadrotor verify the effectiveness of the proposed method.The visual algorithms of the system run in real-time on a onboard minicomputer,without offline processing of the ground station.Real flight results show that the system can achieve visual navigation in large,outdoor and GPS-denied environments.