| 摘要: |
| 视觉感知研究是风景园林学科的重要方向之一。城市河道是建成环境的重要组成部分，对其进行视觉感知研究有利于揭示河道景观特征与公众感知之间
的关系。目前，大数据、深度学习和虚拟现实等技术的快速发展为建成环境视觉感知研究提供了技术支撑。在此背景下，提出基于计算机视觉技术和沉浸式虚拟
现实技术的城市河道景观视觉特征分析方法，以实现对大尺度河道景观灵活且高效的视觉评价。以伦敦泰晤士河为例，利用计算机视觉技术对水面视角的全景图
像进行像素级语义分割，实现多种景观要素的识别与量化分析。同时，通过沉浸式虚拟现实技术获取受试者对城市河道景观的主观视觉感知评价，并进行总体比
较及空间分布特征分析。基于以上分析，探讨客观视觉分析与主观视觉感知结果之间的相关性，并采用随机森林算法分析影响主观视觉感知的客观环境因素。研
究方法弥补了传统定性分析的不足，可为城市河道环境评价和优化提升提供依据，并为相关研究提供新思路 |
| 关键词: 风景园林 河道景观 图像语义分割 虚拟现实 随机森林 视觉感知 |
| DOI：10.19775/j.cla.2025.02.0078 |
| 投稿时间：2023-11-30修订日期：2024-06-17 |
| 基金项目:浙江农林大学科研启动基金(2023LFR076) |
|
| Urban Riverscapes Visual Perception Based on Computer Vision |
| LUO Junjie,LEI Zexin,KONG Xinya,XU Wenhui* |
| Abstract: |
| Visual perception, as a crucial component of landscape environment evaluation,
plays a essential role in assessing urban environmental quality and offers scientific insights
that support urban planning and landscape design. Humans primarily interact with their
surroundings through visual input, making it a vital factor in comprehending and designing
spaces that resonate with public needs and preferences. Previous studies have extensively
examined how visual elements in built environments, such as streets and urban squares,
influence public perception, focusing on factors like green visibility, openness, and spatial
configuration. These studies provide valuable insights into how urban environments
evoke feelings of comfort, safety, and aesthetic appeal. However, research on the visual
perception of urban rivers and waterside routes, which present unique characteristics and
experiential potential, has remained relatively unexplored. Urban rivers are not only vital
natural components of the urban landscape but also offer recreational, tourism, and fitness
functions that add significant social and ecological value. The unique perspective and
experiential demands provided by water-based views enhance the importance of waterside
routes in the development and management of urban river spaces. As cities increasingly
incorporate waterfronts as focal points of urban design, understanding the impact of the
visual characteristics of riverscapes from the water surface perspective on public perception
has considerable practical value. In particular, water-based perspectives allow individuals
to experience the urban environment dynamically, creating novel visual sequences and
sightlines that differ from conventional land-based views. This necessitates specialized
assessment techniques that can accurately capture and quantify the characteristics of urban
river landscapes. To address this gap, this study employs cutting-edge computer vision
and immersive virtual reality (IVR) technologies to perform an in-depth visual perception
analysis of the urban river landscape along London's River Thames. This approach aims
to provide an efficient, objective, and comprehensive assessment of river landscapes by
integrating advanced technological tools. The data source for this study includes waterside
panoramic images acquired from Google Maps. Using the Google Maps API, 100
panoramic image points were collected across a 5-kilometer study area, yielding robust
datasets for subsequent analysis and forming a foundation for systematic evaluation.
Through semantic segmentation with the SegFormer model, pixel-level identification and
quantification of landscape elements were achieved, enabling the precise categorization
and measurement of visual features. Based on previous research and the characteristics
of riverside environments, 6 objective visual evaluation metrics were selected: building
visibility, green visibility, sky visibility, hard shoreline visibility, dynamic object visibility,
and water visibility. Each of these metrics quantitatively represents a distinct spatial
characteristic of the urban river landscape, providing a comprehensive understanding of
how visual features are distributed across the study area. This distribution was further
normalized to facilitate intuitive analysis of visual feature trends, highlighting variations
across different sections of the river and enabling researchers to identify patterns in
landscape composition. In terms of subjective visual perception, this study utilized
the PICO NEO 3 VR headset, which facilitated an immersive experience by allowing
participants to experience the Thames landscape virtually. To capture participants'
subjective evaluations, 4 perception indicators were selected based on the MIT Place Pulse
data set: sense of affluence, vitality, boredom, and scenic beauty. These indicators were
chosen to capture a holistic view of public perception, evaluating the Thames landscape's
quality from economic perception, spatial vitality, and aesthetic experience perspectives.
A sample of 120 participants provided subjective feedback on various river landscape
areas, contributing valuable data for the analysis of how landscape element configurations
influence subjective perception outcomes. By integrating visual semantic segmentation
data with subjective evaluation data, this study yielded quantitative insights into the visual
features of different regions along the Thames and their distribution patterns within public
perception. Our findings on objective visual features revealed several notable correlations.
For example, a significant positive correlation was found between green visibility and
vitality, suggesting that green spaces positively influence public perceptions of energy
and vibrancy. Additionally, building visibility showed a positive correlation with a sense
of affluence in certain regions, highlighting how architectural elements along the river
contribute to an impression of prosperity. This highlights the impact of riverside landscape
composition on public perception and suggests potential strategies for enhancing urban
river environments. Furthermore, The presence of dynamic elements such as boats and
cars exhibited a significant negative correlation with boredom, indicating that visual
dynamics contribute to feelings of vitality and reduce perceptions of monotony. Spatial
analysis revealed that green visibility was higher in central river areas, while hard
shoreline visibility was more prominent in upstream and downstream regions. This finding
aligns with the public perceptions of safety, as hard shorelines are often associated with
structural integrity and accessibility, influencing the sense of security along different parts
of the river. Through the innovative integration of computer vision semantic segmentation
and immersive virtual reality, this study achieved a comprehensive evaluation of diverse
visual features and subjective perceptions of urban river landscapes from a water-based
perspective. The findings emphasize that objective visual features: such as green visibility,
building visibility, and dynamic object visibility-significantly influence public perception,
offering a scientific basis for urban river landscape design and management. This research
highlights how green spaces, architectural presence, and visual dynamism contribute to
perceptions of vitality, affluence, and safety, emphasizing the importance of a balanced
landscape composition for positive public engagement. The methodological innovation of
integrating computer vision and IVR technologies not only enhances the efficiency of data
acquisition and analysis but also offers a replicable technical approach for future waterbased
landscape perception studies. By capturing and analyzing public perception data
through immersive technology, this study sets a precedent for efficient, scalable evaluations
of urban waterscapes. Future research can build upon these findings by exploring the
visual perception characteristics of other urban riverscapes, and expanding data collection
to a global scale to validate patterns across diverse cultural and geographical contexts.
Such research would further refine our understanding of public perception dynamics in
waterfront environments, fostering more targeted and responsive urban landscape designs
that cater to the needs and preferences of urban residents. |
| Key words: landscape architecture riverscapes imagery semantic segmentation virtual reality random forest visual perception |
