ARTIFICIAL INTELLIGENCE BASED SOLUTIONS FOR URBAN TRANSPORTATION SYSTEMS

Abstract

In urban transportation systems, traffic monitoring and management can suggest better routes to the travelers, reduce travel time and fuel consumption, increase road capacity, and passenger safety. Such transportation applications require huge amount of dataset to analyze and derive urban traffic situation. Recent innovations in communication and information technology have opened various sources of data generation. For example, different social media platforms and Traffic sensors made it possible to collect a large dataset and process it to improve the urban transportation systems. These large collected datasets give rise to the need for techniques to discover and understand hidden patterns. These discovered trends and patterns can further be used in various decision making. With the advent of such massive data collection ways, urban transportation systems have been changed significantly into more powerful Artificial Intelligence (AI) algorithm driven systems with optimized performance. Data-driven AI-based learning algorithms employ large datasets from different sources and can be processed into patterns necessary for transport contributors. However, larger datasets come with new challenges in data processing steps. Further, With the expansion of Intelligent Transportation Systems (ITS), it is becoming more challenging to accurately predict urban transport systems and analyze city traffic. The decision ability of machines powered by Artificial Intelligence and Machine learning algorithms has made a paradigm shift in modern society. This thesis proposes different machine learning (ML) algorithms considering urban transportation characteristics, useful for solving urban transportation problems utilizing unstructured text-based social media datasets and numerical traffic sensor datasets. Primarily, this work addresses the issues like accurately detecting traffic information and analyzing the stakeholder’s sentiments from the Twitter-based social media datasets and precise Abstract traffic information utilizing traffic sensor data. A few of the urban transportation challenges proposed in this work are discussed further. With the rapid urbanization and exponential increase of motorized vehicles, frequently occurring traffic events like accidents, potholes, traffic-jam, road maintenance etc., affect daily life traveling. Confined use of road sensors limits the effectiveness of such traffic disturbing event detection. In this context, social media platforms and microblogging websites like Facebook, Twitter, etc. are becoming popular among the people to share the events and things which affect their daily life. In our first objective, an integrated methodology is suggested that uses machine learning (ML) models to detect the traffic events from usergenerated social media data. In order to collect the traffic related tweets, a dictionary of traffic-related keywords has been formed. The novel combinatorial feature generation approach (CFGA) is the main contribution of this work. The proposed CFGA uncovers appropriate associations among the keywords of tweet and extracts the correlated keywords from the collected data. Such keywords are denoted as set phrase. The set phrases may comprise of single or multiple words of a tweet. These set phrases may be used as keywords for event-related data collection for further analysis. The frequently occurring set phrases are identified using the notion of support, which signifies the percentage of tweets containing relevant keywords. Since the nature of different events may also be different, therefore, a hard-coded value for support threshold will not be beneficial. A hyper-parameter named as support (!) is tuned for finding threshold value that is used to obtain the set phrase. This process sets up a database of frequently occurring set phrases that can signalize traffic-related events. This database of set phrases are then used as input for ML-based classifier to identify the text that contains traffic related events. The results of the proposed approach depict that if suitable support is chosen, then proposed CFGA increases the accuracy of supervised classification models for extracting traffic information from twitter data. Transportation is an essential part of human life, therefore, it is important to understand the opinion of the general public in different aspects of transportation. Furthermore, the citizens’ opinion is the actual customer-based performance metric which can be used to assign priorities to the urban traffic issues. Further, Social media is the cost effective source of large amount of dataset, which reflects sentiments and provides the possibility of opinion vi Abstract mining in this context. This work of the thesis proposes an opinion mining approach based on traffic-related tweets to find the citizens’ sentiment for urban transportation issues. In order to show the prevalence of transportation on social media, the location-based traffic related tweets, written by individuals expressing their sentiments about different transport services have been mined, preprocessed, and then a dictionary-based approach is used for the calculation of sentiment and classification of sentiment polarity to evaluate the satisfaction of transportation users. ML covers the biggest part of artificial intelligence and is used extensively in transportation systems from travel prediction to route choice modeling. The third work of the thesis utilizes ML-based techniques to obtain more understanding about urban traffic patterns by analyzing hourly and daily variation in urban traffic flow dataset. A model has been developed for the analysis of spatial and temporal patterns in urban traffic data. Model development involves the formulation of algorithms to be applied to the data and choice of various metrics to evaluate the ML algorithms. Final results of the work are analyzed to determine the various factors that affect the traffic flow patterns in an urban area. The sensor-based average speed estimation of vehicles on road segments is desirable while determining the route of choice. Such estimation activity requires large traffic data to develop an effective traffic prediction model. The sensors deployed on road networks collect traffic data continuously. This results into the generation of large time series of traffic data. Accurately capturing the temporal dynamics of such time-series traffic data with improved prediction performance is an open challenge. The fourth work of the thesis proposes a solution to the problem of route choice behaviors and traffic congestion. An innovative travel speed prediction model using decomposition based smooth time series has been proposed for non-stationary and non-linear temporal dynamics of traffic time series data. The decomposition of time series data is aimed to capture the temporal characteristics of traffic data by improved data representation. The purpose of smoothing is to filter out the irrelevant and excessive fluctuations in time-series data, which may lead to improved model complexity and performance. This way, the thesis proposes solutions to the urban transportation related problems of accurately predicting traffic information using a text-based social media dataset and numerivii Abstract cal traffic sensor dataset. The work promotes the applications of ML algorithms in intelligent urban transportation systems. The proposed solutions of the thesis are tested with their existing counterparts. In the first and second problems, Twitter based social media data is used to develop the model, and precision, recall, and F-measure are used to evaluate the model’s performance. Whereas numerical traffic data is used in third and fourth problems and Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE) error is considered for evaluating the model’s performance in fourth work.