Research On Key Technologies Of Crowd Trajectory Data Mining For Urban Planning

The urbanization of China is in the process of long-term and high-speed advancement.Efficient yet reasonable urban planning work is important to ensure the high-quality urbanization progress.However,there are still challenges in urban planning.Firstly,the planning work relies on the perception of the current situation of the city and the collection of urban information,while the acquisition of these information mainly relies on manual offline collection,which not only consumes a lot of labor costs,but also affects the planning progress;Secondly,the planning scheme design mainly relies on traditional solutions such as surveys and simple rule-based methods,resulting in the over-utilization/underutilization of planed facilities.The 19 th National Congress of the Communist Party of China clearly stated that it is necessary to "promote the integration of the Internet,big data and artificial intelligence with the real economy to build a smart society",which points out the direction for intelligent urban planning.With the popularization of mobile Internet applications,the trajectory data are generated in urban regions,which perceive the real-time status of the city,and contain the traveling patterns and demands of the crowd,bringing new opportunities for intelligent urban planning.This work focuses on trajectory data to study the two aspects of problems in urban planning,including the urban information perception and planning scheme recommendation.The detailed research works and contributions of this paper are as follows:Firstly,we study the problem of the POI alias discovery using trajectory data,which helps urban planers to better understand the geo-spatial information of the city.We first propose the concept of mobility profile,which effectively depicts the geo-spatial features of the POI entities corresponding to the POI names;Then we propose local region estimation algorithm,by which an alias siamese deep network is designed,which works in local-global paradigm,enhancing the model with effective latent representations for large and sparse mobility profiles,yet keeping the input size small.Our method can provide aliases with high precision,and significantly cut down the data tagging labor.Secondly,we study the problem of illegal lane occupation detection using trajectory data,which helps urban planers to precisely extend facilities to address illegal occupation,and improve the utilization of lands.Traditional methods rely on police patrol as well as low-coverage cameras,which is not suitable for efficient and high-coverage detection.To address this issue,we propose a ubiquitous framework to detect the fine-grained urban anomalies using trajectories.On one hand,we tail the trajectory cleaning and map-matching algorithms for the low-speed bike trajectories with unique spatial-temporal properties,and aggregate the sub-trajectories by matched road segments in batches to capture the subtle impact from fine-grained anomalies;On the other hand,as the labeled data are hard to acquire,we propose a novel one-class classification model for anomaly detection via hypothesis tests on sub-trajectory distribution equality.The proposed method can note only effectively capture the location of the occupations without additional human labor,but also improves the patrol efficiency by a lot.Thirdly,we study the problem of trajectory data-driven network-like facilities plan scheme recommendation.Traditional facility planning collect the demand via inefficient surveys and design the planning scheme by experience or simple rule-based methods.Different from point facilities,we focus on network-like facilities like bike lanes or green belts,and define the problem of trajectory driven network-like facility planning,which is proven to be NP-hard.Inspired by the comprehensive analysis of the characteristics of the trajectory distribution,we propose Spatial Density Clustering Initialization based Sub-Network Expansion algorithm,which is validated effective in the real-world scenario of bike lane planning.To improve the efficiency,we propose the novel TS-index to reduce the time complexity of the impact score gain computation,and design to deploy the algorithm on distributed computing framework for better scalability.Forth,we study the problem of cross-city trajectory generation for mobility distribution deduction.Since that transportation related facility planning relies on the trajectories in the city,while in new cities or cities with poor usage of web applications,the trajectory acquisition is difficult.To this end,we propose cross-city trajectory generation framework,which performs zero-shot trajectory generation in the cities with only spatial information like POIs and road networks,and the generated trajectories take place of true trajectories for urban planning.First,we introduce the concept of mobility intention space.In this space,the Origin-Destination distributions of different cities are similar;Secondly,we validate the applicability of domain generalization on modeling unified the mobility intentions distribution,and thus the mobility intention generation model of the target city is constructed;Finally,based on candidate path generation and preference ranking,an inter-OD path generation model is proposed,which assigns the choice probability of each trajectory.Both the heatmap and the quantified distribution difference of generated trajectories validate the effectiveness of our method on cross-city trajectory generation task.