Object Detection And Obstacle Avoidance System For UAV Power Inspection Based On Deep Learning

Nowadays,with the development of multi-rotor UAVs intelligent system technology,the development of artificial intelligence and pattern recognition technology in UAVs intelligent system is a topic of special concern in the field of science and technology recently.UAVs is widely used in military,agricultural,patrol inspection and other fields due to its low operating cost,quick response and convenient use.UAVs inspection,as a new inspection technology,uses visible light and infrared thermal imager and other inspection equipment to inspect the transmission lines,which has the advantages of high speed,high working efficiency,free from regional influence,high inspection quality and high safety.Among them,the research of machine vision obstacle avoidance technology is particularly important in the intelligent electric patrol inspection of UAVs.In the autonomous visual obstacle avoidance system of UAVs,how to quickly find the flight course of obstacle avoidance,identify the target type of obstacle and prevent collision is a key link to ensure the safety and success of autonomous obstacle avoidance.To achieve the security of the UAV autonomous obstacle avoidance and intelligent,and in the light of the key technologies for UAVs visual obstacle avoidance,based on binocular vision sensor unmanned aircraft structures,deep learning visual processing module,carried out the UAV heading identification and visual collision avoidance obstacle avoidance system of research,the main research content includes the following three aspects:Based on the Inception-Resnetv2 deep learning classification network,an improved Inception-Resnetv2 network is proposed for the course prediction of the UAV visual obstacle avoidance system.Among them,the improved network structure not only ensures the high prediction accuracy but also reduces the computational complexity of the convolutional neural network.Experiments were performed on the Fine-Grain Recognition data set in the Vision group at Oxford University,and the improved network model achieved an average accuracy of 92.50%.The network model is carried in the UAV's onboard vision processor to achieve the prediction of the direction of the electric tower obstacles,with the prediction accuracy up to95.63%.Based on Yolov4-Tiny deep learning detection algorithm and Mobile Net basic network,an improved Yolov4-tiny detection algorithm is proposed to improve the accuracy of target detection.Based on binocular vision sensor information mechanism,this paper puts forward a method of visual inspection of anti-collision efficient performance of the improved Yolov4-tiny detection algorithm,realize the obstacle areas in the back image identification,and through the data in the image preprocessing to remove obstacles redundant information outside the region environment,through the Surf polar constraint matching algorithm for obstacle area feature matching,using the least square method to match the target feature points are obstacles to the three dimensional coordinates,using the depth information for calculating three-dimensional,unmanned aerial vehicles position information and target detection results,draw the anti-collision decision area,The position of the UAV and the current obstacle target is judged to realize the obstacle target detection and visual collision avoidance.Based on DJI F450 frame,Pixhawk flight control,Nvidia TX2 processor,binocular vision and other sensors,an experimental platform for the UAV vision processing was built.Based on deep learning image recognition network,a specific scheme of obstacle avoidance course recognition of the UAV is designed for the power tower obstacles in different scenes in the power patrol task,and the obstacle avoidance course prediction experiment is carried out in the actual power patrol task.Secondly,based on the deep-learning target detection algorithm and binocular vision imaging principle,visual anti-collision experiments were carried out in the outdoor environment against common obstacles in electric power patrol,which verified the practicability and effectiveness of the system.