Target Detection And Visual Positioning System Of UAV Based On ROS

With the continuous development of UAV technology,various industries put forward certain requirements for the autonomous flight technology of UAV,and the level of flight intelligence of UAV will be higher and higher in the future.In these specific areas,the role of UAV plays an irreplaceable role in the critical moment,reducing human cost and improving work efficiency.Deep learning has developed rapidly in computer vision in recent years.At the same time,with the introduction of hardware,it is a trend to process intensive data computing tasks at the edge.Therefore,the UAV is equipped with the embedded GPU platform TX2,using the technology of computer vision,using the method of deep learning to identify and locate the target,and developing a UAV target detection and visual positioning system,which is not only of practical engineering value,but also of social significanceThe purpose of this paper is to design and implement a general system of target detection and visual positioning based on ROS UAV,which is equipped with Airborne Camera zenuse x5s by using industry application machine M210 of Dajiang,equipped with embedded GPU platform TX2 core board kit of NVIDIA as hardware platform,equipped with NVIDIA’s exclusive jetpack4.1 package on TX2,and built the working space of ROS robot operating system.In the working space based on ROS,the required functions of UAV scene application are developedThe main work of this paper includes the following aspects:firstly,build the ROS environment workspace based on TX2 platform,apply the DJI onboard SDK in ROS,and the subsequent work will also be realized in ROS workspace,which is also convenient for compatibility with future function expansion.Then for the real-time detection of the micro object in the UAV scene,choose the yorov3 network model.The network model is improved and optimized for the micro object,and its speed and accuracy are better in the target detection.In the framework of Darknet deep learning,the design,implementation,training and prediction of the network are all under this framework,which realizes the end-to-end training prediction and ensures its accuracy and reliability.Then it uses vision to locate the space position of the UAV itself,using the latest apriltag visual reference library,tag family selects tag36h11,and uses three-layer nesting mode for the UAV scene to realize the precise space positioning of different heights of the UAV.With the accurate position,the UAV in the future will be able to fully realize the automatic operation of autonomous take-off,cruise and landing The UAV can be monitored and its own security can be guaranteed.Finally,the result information of target detection and visual positioning needs to be sent to the mobile terminal in real time for real-time monitoring and data cloud upload.Information return according to the customized communication protocol,the protocol conforms to the mavlink 2 standard,encapsulates the result information data,and sends it to the mobile terminal by using the DJI onboard SDK interface,which is parsed and used by the mobile terminal.The system is developed and simulated on the local general computer,and finally transplanted to TX2.