Visual Drone Pose Estimation And Motion Control

With the rapid development of drone technology and deep learning technology,drones are widely used in various automation technologies.Among them,the pose estimation and motion control of drones are key technologies for drones to achieve automation and are important components for drones to achieve automated tasks.The use of stable and robust positioning technology is crucial for the safe flight of drones.In traditional drone pose estimation and motion control,the drone’s flight control system uses IMU and geomagnetic sensors for attitude calculation and GPS for position calculation.However,this method generally has the problem of being affected by electromagnetic interference in practical applications,which seriously affects the safety of drones in automated tasks.Pose estimation and motion control of drones based on vision can effectively avoid problems caused by external electromagnetic interference.The main innovations and work of this paper are as follows:(1)A method of data augmentation for drone pose estimation data sequence is proposed.In vision-based pose estimation,by masking the sample data to adapt to the pose estimation at different speeds,the speed diversity of the dataset is enhanced,and the robustness of the pose estimation network is improved.In vision-based motion control,the sliding window method is used to generate sub-sequence datasets from data sample sequences to enhance the generalization ability of the motion control network.(2)To address the scale uncertainty and motion magnitude affecting pose estimation accuracy in monocular vision pose estimation,a monocular depth estimation network is incorporated into the pose estimation and motion control network to reduce scale uncertainty.Then,the ACMix self-attention mechanism is added to the feature extraction network to establish long-distance dependencies and better express spatial features.Finally,spatial pyramid pooling is used to enhance the scale information contained in the features.These operations improve the generalization performance of the drone pose estimation and motion control network.(3)A drone platform for visual pose estimation and motion control is built to verify the effectiveness of the algorithm.On the drone platform,RGBD is used to obtain depth ground truth for training the monocular depth estimation network,and high-precision visual TAG is used to obtain the ground truth of the drone’s pose estimation.Experiments show that the pose estimation and motion control algorithms proposed in this paper have reached a basically usable level in real scenarios.