Learning spatiotemporal patterns for monitoring smart cities and infrastructure

Publication Year: 2018 Publication Type : Thesis

Abstract:


Recent advances in the Internet of Things (IoT) have changed the way we interact with the world. The ability to monitor and manage objects in the physical world electronically makes it possible to bring data-driven decision making to new realms of city infrastructure and management. Large volumes of spatiotemporal data have been collected from pervasive sensors in both indoor and outdoor environments, and this data reveals dynamic patterns in cities, infrastructure, and public property. In light of the need for new approaches to analysing such data, in this thesis, we propose present relevant data mining techniques and machine learning approaches to extract knowledge from spatiotemporal data to solve real-world problems.

Many challenges and problems are under-addressed in smart cities and infrastructure monitoring systems such as indoor person identification, evaluation of city regions segmentation with parking events, fine collection from cars in violations, parking occupancy prediction and airport aircraft path map reconstruction. All the above problems are associated with both spatial and temporal information and the accurate pattern recognition of these spatiotemporal data are essential for determining problem solutions. Therefore, how to incorporate spatiotemporal data mining techniques, artificial intelligence approaches and expert knowledge in each specific domain is a common challenge.

In the indoor person identification area, identifying the person accessing a secured room without vision-based or device-based systems is very challenging. In particular, to distinguish time-series patterns on high-dimensional wireless signal channels caused by different activities and people, requires novel time-series data mining approaches. To solve this important problem, we established a device-free system and proposed a two-step solution to identify a person who has accessed a secure area such as an office.

Establishing smart parking systems in cities is a key component of smart cities and infrastructure construction. Many sub-problems such as parking space arrangements, fine collection and parking occupancy prediction are urgent and important for city managers. Arranging parking spaces based on historical data can improve the utilisation rate of parking spaces. To arrange parking spaces based on collected spatiotemporal data requires reasonable region segmentation approaches. Moreover, evaluating parking space grouping results needs to consider the correlation between the spatial and temporal domains since these are heterogeneous. Therefore, we have designed a spatiotemporal data clustering evaluation approach, which exploits the correlation between the spatial domain and the temporal domain. It can evaluate the segmentation results of parking spaces in cities using historical data and similar clustering results that group data consisting of both spatial and temporal domains. For fine collection problem, using the sensor instrumentation installed in parking spaces to detect cars in violation and issue infringement notices in a short time-window to catch these cars in time is significantly difficult. This is because most cars in violation leave within a short period and multiple cars are in violation at the same time. Parking officers need to choose the best route to collect fines from these drivers in the shortest time. Therefore, we proposed a new optimisation problem called the Travelling Officer Problem and a general probability-based model. We succeeded in integrating temporal information and the traditional optimisation algorithm. This model can suggest to parking officers an optimised path that maximise the probability to catch the cars in violation in time. To solve this problem in real-time, we incorporated the model with deep learning methods. We proposed a theoretical approach to solving the traditional orienteering problem with deep learning networks. This approach could improve the efficiency of similar urban computing problems as well. For parking occupancy prediction, a key problem in parking space management is with providing a car parking availability prediction service that can inform car drivers of vacant parking lots before they start their journeys using prediction approaches. We proposed a deep learning-based model to solve this parking occupancy prediction problem using spatiotemporal data analysis techniques. This model can be generalised to other spatiotemporal data prediction problems also.

In the airport aircraft management area, grouping similar spatiotemporal data is widely used in the real world. Determining key features and combining similar data are two key problems in this area. We presented a new framework to group similar spatiotemporal data and construct a road graph with GPS data. We evaluated our framework experimentally using a state-of-the-art test-bed technique and found that it could effectively and efficiently construct and update airport aircraft route map.

In conclusion, the studies in this thesis aimed to discover intrinsic and dynamic patterns from spatiotemporal data and proposed corresponding solutions for real-world smart cities and infrastructures monitoring problems via spatiotemporal pattern analysis and machine learning approaches. We hope this research will inspire the research community to develop more robust and effective approaches to solve existing problems in this area in the future.


BibTex:

@phdthesis{ShaoWei2018Lspf, title = {Learning spatiotemporal patterns for monitoring smart cities and infrastructure},
    abstract = {Recent advances in the Internet of Things (IoT) have changed the way we interact with the world. The ability to monitor and manage objects in the physical world electronically makes it possible to bring data-driven decision making to new realms of city infrastructure and management. Large volumes of spatiotemporal data have been collected from pervasive sensors in both indoor and outdoor environments, and this data reveals dynamic patterns in cities, infrastructure, and public property. In light of the need for new approaches to analysing such data, in this thesis, we propose present relevant data mining techniques and machine learning approaches to extract knowledge from spatiotemporal data to solve real-world problems.

Many challenges and problems are under-addressed in smart cities and infrastructure monitoring systems such as indoor person identification, evaluation of city regions segmentation with parking events, fine collection from cars in violations, parking occupancy prediction and airport aircraft path map reconstruction. All the above problems are associated with both spatial and temporal information and the accurate pattern recognition of these spatiotemporal data are essential for determining problem solutions. Therefore, how to incorporate spatiotemporal data mining techniques, artificial intelligence approaches and expert knowledge in each specific domain is a common challenge.

In the indoor person identification area, identifying the person accessing a secured room without vision-based or device-based systems is very challenging. In particular, to distinguish time-series patterns on high-dimensional wireless signal channels caused by different activities and people, requires novel time-series data mining approaches. To solve this important problem, we established a device-free system and proposed a two-step solution to identify a person who has accessed a secure area such as an office.

Establishing smart parking systems in cities is a key component of smart cities and infrastructure construction. Many sub-problems such as parking space arrangements, fine collection and parking occupancy prediction are urgent and important for city managers. Arranging parking spaces based on historical data can improve the utilisation rate of parking spaces. To arrange parking spaces based on collected spatiotemporal data requires reasonable region segmentation approaches. Moreover, evaluating parking space grouping results needs to consider the correlation between the spatial and temporal domains since these are heterogeneous. Therefore, we have designed a spatiotemporal data clustering evaluation approach, which exploits the correlation between the spatial domain and the temporal domain. It can evaluate the segmentation results of parking spaces in cities using historical data and similar clustering results that group data consisting of both spatial and temporal domains. For fine collection problem, using the sensor instrumentation installed in parking spaces to detect cars in violation and issue infringement notices in a short time-window to catch these cars in time is significantly difficult. This is because most cars in violation leave within a short period and multiple cars are in violation at the same time. Parking officers need to choose the best route to collect fines from these drivers in the shortest time. Therefore, we proposed a new optimisation problem called the Travelling Officer Problem and a general probability-based model. We succeeded in integrating temporal information and the traditional optimisation algorithm. This model can suggest to parking officers an optimised path that maximise the probability to catch the cars in violation in time. To solve this problem in real-time, we incorporated the model with deep learning methods. We proposed a theoretical approach to solving the traditional orienteering problem with deep learning networks. This approach could improve the efficiency of similar urban computing problems as well. For parking occupancy prediction, a key problem in parking space management is with providing a car parking availability prediction service that can inform car drivers of vacant parking lots before they start their journeys using prediction approaches. We proposed a deep learning-based model to solve this parking occupancy prediction problem using spatiotemporal data analysis techniques. This model can be generalised to other spatiotemporal data prediction problems also.

In the airport aircraft management area, grouping similar spatiotemporal data is widely used in the real world. Determining key features and combining similar data are two key problems in this area. We presented a new framework to group similar spatiotemporal data and construct a road graph with GPS data. We evaluated our framework experimentally using a state-of-the-art test-bed technique and found that it could effectively and efficiently construct and update airport aircraft route map.

In conclusion, the studies in this thesis aimed to discover intrinsic and dynamic patterns from spatiotemporal data and proposed corresponding solutions for real-world smart cities and infrastructures monitoring problems via spatiotemporal pattern analysis and machine learning approaches. We hope this research will inspire the research community to develop more robust and effective approaches to solve existing problems in this area in the future.
},
   
    author = {Shao, Wei},
    keywords = {Smart city;Device-free;Pattern recognition;Spatiotemporal data;Parking;Trajectory},
    language = {eng},
    school = {RMIT University},
    year = {2018},
   
}

Cite:

Related Publications

RUP: Large Room Utilisation Prediction with carbon dioxide sensor
Type : JournalArticle
Show More
A Scalable Room Occupancy Prediction with Transferable Time Series Decomposition of CO 2 Sensor Data
Type : JournalArticle
Show More
Topical Event Detection on Twitter
Type : ConferenceProceeding
Show More

© 2021 Flora Salim - CRUISE Research Group.