| In recent years,with the rise of deep learning task application in multi-rotor UAV,the research of combineing artificial intelligence technology with pattern recognition technology to realize target detection and automatic obstacle avoidance in intelligent UAV is of scientific significance and practical value.At this stage,our nation is vigorously developing the power industry,and the task of power inspection is becoming gets more and more attention.However,problems of the traditional manual inspection method such as low efficiency,high labor and material cost,and low safety problems are also gradually emerging.Therefore,the development of drones to unmanned operations has become a trend.In order to study the autonomous obstacle avoidance of electric inspection UAV in unmanned operation,to research the detection of various obstacles in the inspection environment,and to improve the deep learning algorithm based on binocular vision sensor and the path planning algorithm of UAV to ensure that the unmanned aerial vehicle can better complete the obstacle avoidance task in the electric inspection unmanned operation task,the main research content includes the following three aspects:(1)The theoretical basics of target detection algorithms and the construction of datasets are introduced.In the part of theoretical basic knowledge,the representative network model and prediction process in the target detection algorithm based on deep learning are expounded,and the convolutional neural network in the target detection model is briefly introduced.The data set construction part mainly introduces the types of obstacles encountered by the electric inspection UAV for obstacle avoidance,collects data from multiple aspects and enhances the data to increase the diversity,which makes data preparation for the target detection algorithm of the subsequent electric inspection UAV for obstacle avoidance.(2)An ED-YOLO(Efficient-Depthwise YOLO)network model based on model compression is proposed to achieve target detection of UAV obstacles.The target detection algorithm is based on Yolov4,which firstly adds a channel attention mechanism to the backbone network to improve detection accuracy without increasing the amount of computation.Secondly,the depth separable convolution is used to replace the traditional convolution in the feature pyramid part to reduce the amount of convolutional computation.Finally,the model compression strategy is used to trim the redundant channels in the network to reduce the model size and improve the model detection speed.Tests were conducted on the dataset independently constructed with 9,600 flight obstacles of power inspection UAV,the obstacle target average detection accuracy for ED-YOLO is reduced only by 1.4% compared with that for Yolov4,while the model size is reduced by 94.9%,the amount of floating point operations is reduced by 82.1%and the prediction speed is increased by 2.3 times.Finally,compared with other existing methods,the proposed ED-YOLO target detection algorithm has the advantages of high precision,small size and fast detection speed,which can well realize the obstacle detection of UAV inspection,and provides a basis for the obstacle avoidance decision of electric inspection UAV.(3)UAV obstacle avoidance path planning in electric power inspection environment is studied.Firstly,the general framework of UAV system and the obstacle avoidance strategy of UAV are proposed.ZED camera is used to build a binocular vision system to obtain three-dimensional spatial information of obstacles,and NVIDIA embedded computer are used to provide computing power for target detection algorithm and binocular vision system.Secondly,the traditional artificial potential field method is introduced,and the defects of the algorithm in the electric power inspection environment are analyzed and explained,meanwhile,the solution is proposed and its effectiveness is demonstrated by simulation and comparison test in MATLAB software.Finally,experimental verification is carried out in outdoor environment. |
PDF Full Text Request