Vehicular Mobility Analysis Leveraging Historical Spatio - Temporal Data

Abstract

The intelligent and efficient urban transportation system of a country is considered a mark of its progress. It plays a backbone role in a country’s development by facilitating trade between regions, reducing travel time costs, and improving accessibility. A gradual increase in vehicles due to the population explosion engenders several issues in the urban transportation system, such as traffic congestion, pollution, and integration among different modes of transportation. To address these issues, analyzing vehicular mobility becomes an essential aspect of understanding and improving the transportation system. Vehicular mobility analysis involves collecting, processing, and analyzing large datasets related to various aspects of transportation systems to identify patterns and trends in data. Many techniques are developed to analyze vehicular mobility using various data sources from domains, including mobile phone data, taxi trajectories data, and social media data. The insufficiency of data mining techniques for handling and processing mobility data, the lack of integrating heterogeneous data, and the lack of an intelligent transportation system are a few important issues that make vehicular mobility analysis challenging in many scenarios. The vehicular mobility analysis broadly involves data modeling and data analysis. Five major contributions are made under these sub-phases in this thesis to address the issue of the lack of data mining techniques for handling, processing, and analyzing trajectory datasets, the integration of heterogeneous data, and the lack of intelligent systems in vehicular mobility analysis. The data modeling phase ensures an efficient representation of the trajectory data for accurate analysis and effective decision-making in vehicular mobility analysis. In the first contribution of this thesis, a new segmentation method is proposed using bearing measurement for trajectory data. The proposed segmentation eliminates multiple waypoints localized over a straight segment(road) to represent it efficiently. The data analysis phase of vehicular mobility analysis involves examining and interpreting the preprocessed data to gain insights into vehicle movement patterns, behaviors, and characteristics in a given area. In this direction, the second contribution proposes an evolutionary multi-objective based clustering algorithm called CLUSTMOSA in the first chapter of the thesis. It is proposed to group similar trajectories, exploiting the proposed segmentation technique for traffic monitoring and analysis. The proposed segmentation technique is further utilized in various data analysis tasks in other chapters also. The third contribution of this thesis proposes another data analysis task, which involves identifying the modes of transportation. It is helpful to extract information on travel time on various routes using different modes. In the proposed work, various GPS point-level characteristics, such as speed, acceleration/deceleration, jerk, and bearing angle, along with a generative model, are exploited for detecting modes of transportation. Another vital task in the data analysis phase is outlier detection, which is explored as the fourth contribution in this thesis. In this thesis, outlier detection is explored using supervised and unsupervised ways. The classification-based anomaly detection is proposed using one-shot learning with a temporal feature-based Siamese capsule network for the GPS trajectory dataset. The temporal feature is extracted using a recurrence plot of chaos theory, and the Siamese capsule network is applied for outlier detection in trajectory. It detects with limited labeled trajectory data. Although the proposed supervised method is capable of detecting outliers efficiently, it has a limitation in using labeled data. Therefore, a spatial-temporal aware variational graph auto-encoder (ST-VGAE) neural network is proposed in the unsupervised category. The proposed method involves encoding the spatial-temporal characteristics of vehicle trajectories into a graph structure to capture the dynamics of trajectories. The proposed approach utilizes the power of graph convolutional networks (GCN) and variational autoencoders (VAEs) to seize the spatial and temporal dependencies for identifying outliers in GPS trajectories without any labeled data. Further, vehicle mobility analysis plays a significant role in developing various applications related to transportation systems. The fifth contribution of this thesis proposes a framework that intelligently identifies accidental hotspots, profiles driving paths, and recommends safer driving locations. It analyzes traffic scenarios using crowd-sourced data from X (formerly (2006–23) Twitter) and employs aspect-based keyword generation to enhance the dataset. The framework mines potential hotspots from Tweets and uses the BERT-BiLSTM model for sentiment analysis on accident-related tweets of accidents in a location. It conducts probabilistic modeling to evaluate the risk at each accident hotspot, considering factors such as the intensity of opinions in the tweets, historical accident data, and GPS data characteristics. A general recommendation framework is introduced to suggest the alertness required for a driver while passing through the hotspot or nearby regions. The performances of the proposed techniques are evaluated using multiple performance metrics and datasets of GPS trajectory data. The proposed techniques perform better against various state-of-the-art methods, and experimental results favor incorporating them into the urban transportation system.