| Drones and electric robots are often used to inspect the high-voltage power lines;the use of drones instead of manual hoisting of electric robots on the power lines can save a lot of time and cost;collisions between aerial vehicles and power lines have frequently occurred in recent years,which causes casualties and property losses.Whether for intelligent operations on power lines or for low-altitude aircraft watching out for transmission lines,a system of identifying and positioning transmission lines is required.Traditional vision algorithms have low accuracy in detecting slender transmission lines in complex scenes,and are greatly affected by environmental factors;the existing power line detection algorithms based on deep learning are not efficient.To alleviate these problems,this paper proposes an algorithm for power line recognition and positioning,based on the fully convolutional neural network of the encoding-decoding structure and binocular vision.This algorithm is suitable for running on a mobile terminal.The algorithm is implemented at a hardware platform which is composed of a mobile GPU and a binocular camera.The main research content of the article is as follows:(1)A convolutional neural network,named Power Line Net,is proposed,which is suitable for mobile terminal operation with encoding-decoding structure.The overall network structure of the model is streamlined and the number of parameters is small;the up-sampling method and loss function suitable for power line segmentation are adopted;an attention module is introduced to further improve the power line recognition performance of the model with little impact on efficiency.In addition,a transmission line data set with complex background and different light conditions is constructed,and a public transmission line data set is relabeled to verify the generalization performance of the model.Compared with the existing power line recognition model,Power Line Net has reached 8.0FPS recognition speed on Jetson TX2,which is doubled;it has better performance on different segmentation evaluation indicators as well.(2)Because the width of power line in the image is small with only 5-10 pixels,it is difficult for the traditional binocular vision feature matching algorithm to extract the features on the power line to calculate the parallax.In response to this problem,the power line detection is firstly performed on both the left and right views,and then the image post-processing is performed to extract the power line contour,and finally the power line is located according to the average parallax of pixels on the left and right view contours.(3)A set of transmission line identification and positioning software system is developed.The system integrates the Power Line Net model and binocular vision algorithm,and has functions such as power line detection and power line positioning.Finally,the various functions of the system are verified through experiments,proving the application value of the system. |
