Research On Real-time Sense And Avoid For Multi-rotor Uavs In Unknown Environment

In recent years,the application of UAV in military and civilian fields has become increasingly widespread,and the research on UAV systems technology has been developed rapidly.However,existing methods based on pre-planning and program control cannot guarantee the flight safety of the UAV.In view of the complex and highly dynamic mission environments and frequent UAV collisions,the real-time perception and autonomous obstacle avoidance have become one of the key challenges for the development of UAVs.Considering the limited sensing ability of airborne sensors,a small rotorcraft UAV with depth camera is taken as the research object to study the realtime perception and autonomous obstacle avoidance of UAV in an unknown environment.The main work of the thesis includes:(1)Aiming at the real-time perception problem of UAV in unknown environment,proposing a local map construction method based on depth image.Firstly,the depth image obtained by the airborne depth camera is transformed into point cloud,and the point cloud is filtered to remove noise points.Taking advantage of the compressibility and dynamic updating of octree,use point cloud and octree structure increment to construct octree map.Combined with ring buffer and octree map,a local occupation map centered on UAV and updated with the flight of UAV was constructed to provide bounded map for the subsequent trajectory planning to realize collision detection.(2)Aiming at the problem of autonomous obstacle avoidance under the condition of UAV local sensing,proposing a real-time trajectory re-planning obstacle avoidance method based on local map.Firstly,the trajectory re-planning problem was decoupled to path search and trajectory fitting,and the trajectory was represented by b-spline to obtain a smooth trajectory that could be followed by UAV.However,the trajectory obtained based on the above method is not the optimal trajectory.Therefore,this paper further proposes a trajectory re-planning method based on b-spline control point optimization.By designing the constraint conditions of the collision-free problem and optimizing the solution,this method generates the optimal collision avoidance trajectory in real time with the flight of UAV.Simulation experiments show that both the b-spline optimization method and the path-search method can achieve rapid avoidance of static obstacles,but the former needs much less time to re-plan the obstacle avoidance track than the latter,and can achieve effective avoidance of dynamic obstacles.(3)Aiming at the discrete/continuous obstacle avoidance action decision problem under the unexperienced condition of UAV,a passive obstacle avoidance method based on DDQN(Double Deep Q-learning Network)algorithm and PPO(Proximal Policy Optimization)algorithm is proposed.Firstly,the reactive autonomous obstacle avoidance problem of UAV is described as markov decision-making process.Then,the mapping relationship between visual perception and action of UAV is established.The forwardlooking depth image,target information and the movement information of the last moment of UAV are taken as the input of deep neural network,and the output is the motion control command of UAV.By means of the interaction between UAV and environment,the training of deep reinforcement learning neural network is completed to realize the end-to-end control of autonomous obstacle avoidance of UAV.Simulation test results show that the proposed two methods can successfully avoid obstacles in the environment and reach the expected destination quickly without initial guidance path and map,which verifies the effectiveness of the proposed methods.However,the convergence speed of PPO algorithm is slightly faster than DDQN algorithm.