A Research On Data-driven Urban Landscape Perception System

With the rapid development of computer vision,machine learning and other related technologies,the research of using urban street view images to obtain large-scale and multi-dimensional quantitative perception of city is becoming more and more active.The existing related studies show that it is reasonable and feasible to use street view images,limited image evaluation data and machine learning algorithm to realize urban multi-dimensional perception.This thesis aims to obtain the perception of Chengdu’s regional beauty and vitality,focusing on the integration of two types of data quantitative scoring model,automatic scoring model based on support vector regression and the main determinants of human perception of urban street view,the main research content is divided into four parts.Data collection and analysis.According to an unified image collection standard,3672 street view images from the main urban area of Chengdu were collected,which were divided into training dataset and extended dataset.6870 rating results and 3360 pairwise comparison results were collected from 31 raters.This thesis analyzes raters’ evaluation habit and quality by using the statistical theory of rater’s internal reliability and rater’s evaluation quality.Quantitative evaluation of street view images.We proposed a quantitative scoring model,which is able to integrate rating data and pairwise comparison data.Due to the differences in the evaluation quality of the raters,different weights are given to each evaluation sample in the image quantitative scoring to optimize the scoring model and improve the accuracy of the scoring model in fitting the real level of the image.The performance of the optimization model is verified by comparing the conflict rate between the model computing results and the contrast pairwise comparison data samples.Automatic quantitative perception of street view images.The support vector regression is used to construct the automatic scoring model of street view images.The features of street view images in training set are extracted as input,and the quantitative scoring value of images is used as output to train the automatic scoring model.We use the coefficient of determination and cross validation to evaluate the generalization effect of the regression model.By comparing the generalization effect of various image features,the optimal feature combination of the two regression models is obtained to evaluate the beauty and vitality of street view images.Finally,the quantitative scores of all the collected images are obtained,and the perception distribution map of human beauty and vitality in Chengdu City is drawn.Analysis and verification of perceptual elements.Through questionnaire survey,we collected 52 data to judge the beauty and vitality of street view pictures.The most important elements of street view were obtained.The proportion of pixels or the number of instances of street view components are extracted by image segmentation,and the correlation analysis is made with the activity quantitative score of corresponding pictures.The final results show that the activity score of street view pictures is positively correlated with the traffic volume,traffic volume and greening of the top three in the statistical results,and the correlation is low with houses and roads.