Speaker
Description
“Healthy Streets” have emerged as a key strategy for enhancing street quality and promoting public health. Within the broad framework of Healthy Streets, this study focuses on their healing qualities and emphasizes healing environments that support attention restoration and emotional recovery. The environments shall improve users' physical, cognitive, and behavioral health. By utilizing streetscape images and deep-learning-based semantic segmentation models, multiple spatial elements are efficiently extracted and quantified for evaluation metrics. However, existing research primarily focuses on the correlation between spatial elements and healing effects, lacking a systematic exploration of the entire process from perception and evaluation to optimization. Therefore, this study employs LoRA fine-tuning technology in AIGC to develop an iterative path for street quality improvement, aiming to explore how AI technologies can facilitate intelligent and dynamic optimization in Healthy Streets studies.
This study was conducted in the research area of the central region of Haining City, Zhejiang Province, China, covering 52.4 square kilometers. The study area includes diverse street types and historical and cultural blocks with favorable walking conditions. AI empowerment was reflected in the whole research design and tested in the study area. Streetscape images were collected at 100-meter intervals along the road network. After semantic segmentation of the images with a CNN model, eight metrics—facade diversity, color richness, building enclosure, street height/width ratio, sky openness, green view index, road width, and relative walking width—were calculated across four dimensions regarding user perception (vitality, comfort, nature, and safety). All images were clustered based on index scores and a stratified sample of representative images was selected for the Elo rating system (ELO) scoring to determine the healing scores. An artificial neural network (ANN) model was trained to assess the overall street healing quality, complemented by multiple linear regression (MLR) to analyze the correlation between healing quality and each spatial element. Subsequently, a LoRA fine-tuning model was trained to identify and generate high-healing-level streetscape images. Leveraging the Stable Diffusion model, streetscape images of historical and cultural blocks were optimally redesigned based on the MLR results, and the ANN model was re-applied to evaluate the enhancements in street healing quality across the four user-perception dimensions.
We completed the acquisition and semantic segmentation of 1,851 streetscape images from the observation points throughout central Haining City. A subsequent cluster analysis resulted in four distinct street types based on their characteristics. To ensure robust scoring validity, 200 representative images as training data underwent over 2,500 rounds of ELO scoring by professionals and students. The ANN model predicted the healing quality score with a reasonable accuracy of 76%. The model was then applied to rank the healing qualities of the remaining 1,651 images, with spatial distribution visualized and analyzed using ArcGIS software. Our MLR analysis revealed that ‘nature’ had the most significant positive impact on healing qualities. Excessive building enclosure in the ‘comfort’ dimension weakened the healing effect. ‘Vitality’ had a limited but positive contribution, and higher motorization levels in the ‘safety’ dimension were linked to noticeable reductions in healing qualities.
This paper innovatively integrates AI technologies into a study of Healthy Streets, establishing an iterative path of “intelligent perception—spatial quality evaluation—rapid response optimization”. In the perception stage, visual elements were automatically extracted using CNN-based semantic segmentation. During the evaluation stage, a trained ANN model outperformed traditional machine learning models in both accuracy and efficiency. In the optimization stage, AIGC tools (a LoRA and a Stable Diffusion model) were employed for local redrawing of streetscapes, enabling rapid design iterations and dynamic optimization. The preliminary results demonstrate that AI technologies hold substantial potential for improving street healing quality, particularly regarding visual aesthetics and emotional restoration.
References
[1] Ulrich, R.S. (1979) 'Visual landscapes and psychological well-being', Landscape Research, 4(1), pp. 17-23.
[2] Wijnands, J.S., Nice, K.A., Thompson, J., Zhao, H. and Stevenson, M. (2019) 'Streetscape augmentation using generative adversarial networks: insights related to health and wellbeing', Sustainable Cities and Society, 49, 101602.
[3] Long, Y. and Tang, J.X. (2019) 'Research progress on large-scale quantitative measurement of urban street spatial quality', Urban Planning, (6), pp. 107-114.
[4] Ye, Y., Zhang, Z.X., Zhang, X.H. and Zeng, W. (2019) 'Measurement of street space quality at human scale: A large-scale and high-precision evaluation framework combining street view data and new analysis techniques', Urban Planning International, (1), pp. 18-27.
[5] Chen, J.J., Zhang, Z.X. and Long, Y. (2020) 'Strategies for improving street space quality oriented towards public health: From the perspective of spatial disorder', Urban Planning, 44(9), pp. 35-47.
[6] Xu, L.Q. and Hu, Y.Z. (2020) 'Healing streets: A new model of healthy streets', Time Architecture, (5), pp. 33-41.
[7] Quintana, M., Gu, Y. and Biljecki, F. (2024) 'My street is better than your street: Towards data-driven urban planning with visual perception', in Proceedings of the 11th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation.
| Keywords | Healthy Streets; Healing Environment; Streetscape Analysis; Deep Learning; AIGC |
|---|---|
| Best Congress Paper Award | Yes |
