With more than half of the world’s population living in urban areas, a huge in-town travel demand is resulted and forecasted to continuously increase in next few decades. Particularly, walking is a natural mode of mobility, which brings substantial socioeconomic benefits to a society. Hence, creating a pleasant walking environment is essential to encourage citizens to walk. Governments in different regions worldwide have been actively setting out visions to incorporate urban walkability into new area development and urban renewal. Therefore, systematic methodologies for walkability monitoring are highly demanded. Existing practices mostly rely on statutory design guidelines, walking facility audits and pedestrian surveys. However, these approaches tend to be time-consuming and labor-intensive, mainly due to the complexity and regional uniqueness of walkability. Therefore, more effective methodologies are needed for walkability monitoring. Nowadays, closed-circuit television (CCTV) cameras are commonly adopted, which provide rich visual information about pedestrian appearance and movement. Conventional computer vision techniques have been applied to automatically process CCTV videos for pedestrian flow analytics. Yet, they require strong prior knowledge of specific camera scenes to design hand-crafted feature extraction mechanism, which appears very tedious and not scalable to different cameras. In recent studies, deep learning techniques have been incorporated, which outperform conventional computer vision methods in several image processing tasks. Deep neural networks are highly flexible in representing abundant and complicated image patterns, whose feature extraction mechanism is automatically optimized via a data-driven approach, i.e. directly feeding input images and labeled output, without tediously hand-crafting intermediate features. Hence, deep learning techniques demonstrate competitive performance for video processing. Therefore, this research aims to incorporate deep learning-based computer vision techniques into CCTV analytics for automatically analyzing pedestrian walking behavior, in order to facilitate pedestrian walkability monitoring. Overall, this thesis consists of three parts. Part 1 is the literature review on existing approaches for walkability monitoring and CCTV video analytics. Among previous studies, four research gaps among existing methods are to be addressed: 1) limited robustness for pedestrian tracking due to challenging scene conditions like occlusion, 2) limited in small areal coverage and cannot analyze pedestrian walking behavior across different areas, 3) not scalable when adapting deep learning models to different scenes requiring ineffective collection of massive training data, and 4) lack of a comprehensive framework to capture infrastructural and pedestrian-related attributes for quantitative walkability analyses. Therefore, Part 2 focuses on extracting pedestrian movement based on automated CCTV analytics, including three works: 1) pedestrian trajectory tracking within single camera, 2) pedestrian re-identification across multiple cameras, and 3) transfer learning for scene-adaptive model adaptation. Subsequently, Part 3 focuses on the development of an integrated framework for walkability analyses integrating infrastructural characteristics and pedestrian behavior with CCTV analytics. More specific elaboration of each research work is as follows. Work 2.1: Pedestrian trajectory tracking within single camera. Existing methods of pedestrian tracking suffer from several challenging conditions, such as inter-person occlusion and appearance variations, which leads to ambiguous identities and hence inaccurate pedestrian flow statistics. Therefore, a more robust methodology of pedestrian tracking is proposed which 1) incorporates high-level pedestrian attributes into enhancing pedestrian tracking, 2) presents a similarity measure integrating multiple cues for identity matching, and 3) includes a probation mechanism for more robust identity matching. Work 2.2: Pedestrian re-identification across multiple cameras. Existing studies of pedestrian flow analytics are limited in a small areal scale, while tracking the movement of a person may require accurate identity matching across different cameras, which is challenging due to the appearance variation and ambiguity among different persons. Therefore, a more robust methodology of pedestrian re-identification is developed which 1) presents a new convolutional neural network architecture that extracts discriminative-and-distributed human features, 2) presents a generic approach of explainable model design by intuitively visualizing feature extraction mechanism of deep learning models, and 3) includes an incremental feature aggregation strategy designed for more robust identity matching. Work 2.3: Scene-adaptive model deployment with transfer learning. Existing deep learning models require massive amount of scene-specific training data, which hinders their scalability of being deployed to practical multi-camera analytical system. Moreover, further optimization of existing training strategies of pedestrian re-identification models is needed, to facilitate their feature extraction performance. Therefore, a more robust framework is developed to enhance pedestrian re-identification which 1) presents a new loss function named similarity loss designed to generically facilitate feature extraction, 2) presents another loss function named similarity distillation loss designed to transfer knowledge across training datasets for reduced labeling effort, and 3) proposes a workflow for what-if facility design evaluation by integrating pedestrian movement behavior with geometric layout. Work 2.4: Integrated framework for walkability evaluation. Previous studies of walkability analyses utilized Building Information Modeling (BIM) and Geographic Information System (GIS) to model infrastructural characteristics, such as the geometric connectivity of paths. Yet, they lack a comprehensive data schema that formalizes the walkability-related attributes for analyses among different pedestrian groups and scenarios. The pedestrian movement behavior is also not fully integrated with infrastructural models, and a walkability scoring mechanism is also needed to quantitatively analyze walkability of different routes. Therefore, a framework for walkability evaluation is proposed which 1) extends the openBIM data schema to formalize the walkability attributes for modeling individual building, 2) extends the openGIS data schema to define the attributes for constructing walkability network by integrating BIM, GIS and pedestrian flow data, and 3) presents a mechanism of walkability scoring to quantify the walking costs of a route considering different pedestrian groups and walking directions. Compared to existing approaches of walkability monitoring, the developed methodologies of CCTV analytics extract pedestrian flow statistics more accurately, with more robust performance when deployed to different scenes and cameras in practical multi-camera analytical systems. The methodologies are generic and also provide insight for improving deep learning models in future research that extract pedestrian movement behavior with CCTV analytics. Furthermore, the proposed BIM-GIS framework formalizes the data requirement and modeling methodology of infrastructural attributes, which provide a basis for future walkability studies when modeling as-built environment. The developed walkability scoring mechanism also provide a basis for future studies to quantitatively analyze the walkability of different areas.
| Date of Award | 2022 |
|---|
| Original language | English |
|---|
| Awarding Institution | - The Hong Kong University of Science and Technology
|
|---|
| Supervisor | Jack Chin Pang CHENG (Supervisor) |
|---|