Research On Obstacle Recognition In Lawn Scene Based On Digital Image Processing

Lawn is an indispensable part of urban greening.At present,lawn mowers are still lack of intelligence.With the development of artificial intelligence,vision-based mobile robot navigation has become a hot issue.This paper mainly studies the identification of obstacles in lawn scenes based on digital image processing,including the following contents:(1)Obstacle segmentationFor the two types of obstacles in the Lawn scene,the image segmentation method based on color and texture was studied.In the color segmentation: the statistical method was used to estimate the range of the hue value of the lawn pixel,and the value of the two ends of the hue range was set as the segmentation threshold,thereby realizing the segmentation based on the hue statistical threshold.The distribution characteristics of the R(Red),G(Green),B(Blue)components of the lawn pixels were analyzed experimentally,and the characteristics that the lawn pixels have the highest G component within a certain range were obtained.Compared with the adaptive hue threshold and excess green features(2G-R-B)adaptive threshold segmentation method,the hue statistics threshold segmentation method has the lowest misclassification rate and the most robust to illumination changes.In the Texture segmentation: The process of analyzing texture feature data for selecting the optimal parameters for extracting texture features was designed.The texture feature samples were produced to train the support vector machine,and the obstacle segmentation based on texture features was realized by the idea of image segmentation operation.Finally,the texture segmentation algorithm was compared with other segmentation algorithms based on Canny edge feature and fuzzy clustering,and the superiority of the texture segmentation algorithm was verified.(2)Obstacle recognitionIdentify obstacles using the convolutional neural network(CNN)algorithm.First,an obstacle recognition data set was obtained by data enhancement and division.Through the experiment,the influence of the learning rate on the model training was analyzed,and the initial value range of the learning rate suitable for the research object of this paper was obtained by gradually increasing the learning rate.The performance of fixed learning rate,decay learning rate and periodic learning rate were compared and analyzed.The design and experiment of super parameters such as the number of convolutional neural network layers and the number of convolution kernels were completed.Finally,the generalization ability of the model was improved from the perspective of data and algorithm,and make the model achieve 95.58% recognition accuracy on the validation set.