Large-Scale Street Space Quality Evaluation Based on Deep Learning Over Street View Image

In the quantitative study of cities, the extraction and appropriate evaluation of the space quality information of urban streets can provide great insight and guidance to urban planners to build more livable urban public space, which is also of great significance for urban management. However, the traditional methods, which mostly use the manual statistical investigation to carry on, are difficult to carry out large-scale objective quantification. To tackle this challenge, this paper presents a complete quantitative analysis method for street space quality score based on street view image analysis. Three quantitative indices (i.e. cleanliness, comfort and traffic) for the evaluation of street space qualities are employed in this study as suggested in literature on urban planning. A new deep learning approach, named as Cross-connected CNN + SVR, is proposed to estimate the street space quality score. A new dataset is constructed based on Baidu Street View image for the training and validation of the proposed framework. Experimental results suggested that the three indices used in this paper is able to reflect the street’s objective visual attributes effectively and the proposed CNN + SVR approach has produced insightful results. The proposed approach has been applied to evaluate the street space quality score of the 2nd ring road district of Chengdu, to demonstrate the value and effectiveness of the proposed work for providing data support and analytics support to urban planners.