Research On Algorithms Of Bird Nest Detection On Power Line

Bird's nests detection in power transmission line is one of the key technology of Smart Grid inspection.Some birds are used to build nests in the power transmission line,which may cause some adversely affect or even bring severe damage to power grid.In the past,bird's nests are usually detected by manual inspection,which is very time-consuming and dangerous.Hence,an automatic nests detection system is be required.In this paper,basing on computer vision and machine learning,an automatic nests detection algorithm is proposed,which can be deployed in power transmission line to monitor the nests without manual inspection.In our research,we construct a database,which consists of more than two thousands pictures of bird's nest building in power transmission line.Generally,the images can be divided into two kinds:(1)Simple image,where the bird's nest is clear and obvious;2.Complex image,where the bird's nest is fuzzy or occluded.Different detection schemes for both two kinds of images are designed and compared with each other.The main work of this paper can summarized as:(1)An unsupervised learning based nest detection algorithm is proposed,where the K-Means and Gaussian Mixture Model(GMM)are deployed.The algorithm is testified by using the first kind of images.In this algorithm,a pre-process is designed to remove the disturbance in the background.As for the bird's nest in the image,an Progressive Hough transformation is used to extract the characteristic of lines and a histograms of edges length is designed.By using the characteristic of histograms character,an unsupervised clustering algorithm is deployed basing on Principal Components Analysis(PCA).The unsupervised learning does not need to carry out a large number of samples of the mechanized marking work,do not need to carry out complex training process,and the algorithm is efficient,but the algorithm is not robust.Therefore,three kinds of supervised learning based nest detection scheme are designed and tested by using both two kinds of images.(2)We classify and tag the images from histogram character which is extracted from first kind of picture samples.Then KNN(K-Nearest Neighbor)algorithm is trained by using these histogram character samples.After training process,the algorithm can detect the nest from test samples.(3)The samples of nest image and non-nest image is extracted from both two kinds of images.Classification is achieved based on AdaBoost algorithm and Haar/LBP(Local Binary Pattern)features.Finally,LBP feature is selected and used for training process.The experiment results shows the LBP based detection algorithm can achieve a better performance.(4)By using the local database,deep learning based CaffeNet,which is an important method of Fast R-CNN,can be fine-tuned to an automatic nest detection system.Finally,the accuracy of the deep learning based supervised algorithm achieved 92.46%.Through a large number of experiments,supervised learning scheme is selected to establish the nest detection system.This is because supervised learning scheme achieve the best robust performance with the inference of complex background and complicated light environment.