Research Of Vision-based UAV Target Detection And Tracking And Its Autoland Technique

In recent years,with the rapid development of computer vision technology,remarkable progress has been made in the fields of autonomous driving,robots and unmanned aerial vehicle(UAV).UAVs are widely used in ecological protection,resource exploration,marine environment monitoring,agricultural operations,urban planning,etc,due to their simple structure,low cost,strong maneuverability and vertical take-off and landing.UAVs will play a more important role in people's real life.Therefore,the research of vision-based UAV target detection and tracking and its autonomous landing technique is very important and necessary.The key technologies of this thesis involved includes target detection,target tracking and flight control methods.First,the two branches of the current target detection algorithm based on convolutional neural network,candidate-based detection algorithm and regression based detection algorithm,are compared.Considering the computing performance of the hardware platform,the YOLOv3-tiny network is finally used for implementing the detection task.Then the target tracking algorithm based on correlation filtering is introduced in detail.These tracking algorithms use cyclic shift dense sampling and converts computation to frequency domain to boost operating speed.This thesis mainly introduces the KCF algorithm and its improved algorithms in terms of scale adaptation and features.Two improvements are proposed based on Fast Discriminative Scale Space Tracker(fDSST)including Color Names(CN)feature and adaptive learning rate for model update.The effectiveness of the improvements is verified on open source datasets.Finally,according to the characteristics of the UAV platform,the coordinate conversion and monocular ranging calculation methods are derived.This program utilizes the target detection and target tracking algorithm mentioned above with the control algorithm to achieve the target tracking and autonomous landing.When M100 is in the tracking state,the speed is controlled by the cascade PID,and when M100 is in the falling state,the single-stage PID control works.This article takes the DJI M100 quadrotor UAV with NVIDIA Jetson TX2 and See3CAM_CU55 camera as the hardware platform for verification.The experimental results show that the target detection,tracking and landing schemes designed in this thesis are stable and reliable,and successfully realized the detection and tracking for known targets and the autonomous landing function.