Narrow Gap Traversing And Pose Estimation Of Trapped Persons For Rotorcraft UAVs

Rotorcraft Unmanned Aerial Vehicles(UAVs)are widely used in military,civil,scientific research and other fields due to their high mobility and flexibility.Robust and efficient target perception for rotorcraft UAVs can greatly improve the ability and efficiency of rotorcraft UAV’s task execution.Taking the post-earthquake rescue scenario as an example,small rotorcraft UAVs can detect and traverse through narrow gaps autonomously to replace rescuers entering the building,and then perform target detection and pose estimation of trapped persons so as to deliver supplies and implement other rescue work as soon as possible after the disaster.According to the above research background,this thesis mainly carries out the research on the autonomous narrow gap detection and traversing with small rotorcraft UAVs,as well as the target detection and pose estimation of trapped persons under weak light conditions.After the earthquake,the narrow gaps into the building are usually irregular in shape,so how to use the onboard depth camera to accomplish reliable traversing with a small rotorcraft UAV is the main task in our work.In order to solve the problem of irregular and narrow gap detection,a fast semantic segmentation network based on RGB-D image is presented in this thesis.On the basis of the output results of the network,a multi-label semantic segmentation post-processing method based on double masks is proposed,which overcomes the problem that the single-label-based semantic segmentation post-processing method is prone to pixel misclassification.To obtain the accurate spatial position of the narrow gaps,the RANSAC algorithm is used to fit the point cloud from the depth camera into a 3-D plane,and the projection image of point clouds belonging to the narrow gap on the plane is also calculated.After that,a parallel line sliding search strategy was proposed to complete the judgment of the passability of the narrow gaps and the calculation of the key way-points for the UAV’s traversing.The other key task in our work is to detect trapped persons and estimate their pose efficiently by using thermal infrared images with a small rotorcraft UAV,and then to calculate the 3-D coordinates of these persons using the onboard depth camera so as to deliver the supplies accurately.To improve the real-time performance of human target detection and pose estimation algorithms,this thesis proposes a one-stage human target detection and pose estimation network for thermal infrared images,which can complete both human target detection and pose estimation in one network inference.Due to the limited onboard computing resources,a knowledge distillation strategy is adopted for training a lightweight model so as to further improve the speed of model inference.In addition,after the registration between the depth image and the thermal infrared image,a 3-D coordinate mapping method of the human body’s key points based on the circular region is designed,so that the 3-D coordinates of the body key points can be obtained for performing supply delivering.In order to verify the effectiveness of the proposed algorithms for autonomous gap traversing,trapped person detection and pose estimation for rotorcraft UAVs in this thesis,a quadrotor UAV platform equipped with a depth camera,a thermal infrared camera and an onboard computer is built by ourselves.A series of experiments are carried out in the artificial testing scenarios in our laboratory.Experimental results and data analysis show that the method proposed in this thesis can effectively complete the task of rotorcraft UAV’s narrow gap traversing and pose estimation of trapped persons,thereby providing a feasible technical solution for UAV’s post-earthquake rescue after an earthquake.