Computing and Information Systems - Theses
Permanent URI for this collection
Search Results
1 - 3 of 3
-
ItemAvoiding Bad Traffic: Analyzing the Past, Adapting the Present, and Estimating the Future( 2021)Mobility is essential for modern societies. However, due to the increasing demand for mobility, traffic congestion poses a significant challenge to economic growth and advancement for many cities worldwide. At the same time, the widespread availability of location-aware devices has led to a sharp increase in the amount of traffic data generated, thereby, providing an opportunity for intelligent transportation systems to emerge as one of the main cost-effective methods for traffic congestion mitigation. This boost in traffic data has led to a new generation of live navigation services that depend on traffic estimation to provide up-to-date navigation advice. Intelligent transportation systems increase the utilization of existing infrastructure and support drivers to make better navigation decisions by providing actionable traffic information. However, a fundamental shortcoming of existing intelligent navigation systems is that they do not consider the evolution of traffic and route drivers based on snapshot traffic conditions. This is especially critical in the presence of traffic incidents, where the impact of the incident introduces significant variation in the traffic conditions around the incident as things unfold. This thesis proposes three contributions focusing on traffic estimation and forecasting to help drivers avoid bad traffic, especially around traffic incidents. The first contribution is an automated traffic management service to help drivers avoid traffic events based on analyses of historical trajectory data from other drivers. Users subscribe to the service and, when a traffic event occurs, the service provides advice based on all drivers' actions during a similar traffic event in the past. We present a solution that combines a graph search with a trajectory search to find the fastest path that was taken to avoid a similar event in the past. The intuition behind our solution is that we can avoid a traffic event by following the traces of the best driver from a similar situation in the past. The second contribution is a system that uses real-time traffic information and fast traffic simulations to adapt to traffic incident impact and generate navigation advice. In this work, we use faster than real-time simulations to model the evolution of traffic events and help drivers proactively avoid congestion caused by events. The system can subscribe to real-time traffic information and estimate the traffic conditions using fast simulations without the need for historical data. We evaluate our approach through extensive experiments to test the performance and accuracy, and quality of the navigation advice of our system with real data obtained from TomTom Traffic API. For the third contribution, we propose effective deep learning models for large-scale citywide traffic forecasting. In addressing this problem, our goal is to predict traffic conditions for thousands of sites across the city. Such large-scale predictions can be used by navigation systems to help drivers avoid congestion. We propose a traffic forecasting model based on deep convolutional networks to improve the accuracy of citywide traffic forecasting. Our proposed model uses a hierarchical architecture that captures traffic dynamics at multiple spatial resolutions. We apply a multi-task learning scheme based on this architecture, which trains the model to predict traffic at different resolutions. Our model helps provide a coherent understanding of traffic dynamics by capturing short and long spatial dependencies between different regions in a city. Experimental results on real datasets show that our model can achieve competitive results while being more computationally efficient.
-
ItemEffective Spatial Feature Extraction via Convolutional Neural Networks( 2021)Spatial data analysis has achieved great success in many real-world applications such as region similarity learning, crime prediction, and traffic prediction. In spatial data analysis, spatial features play an important role. In this thesis, we propose novel techniques based on the Convolutional Neural Network (CNN) and its variants to learn spatial features from two important aspects, namely Point of Interest (POI) features and POI relationships. We focus on the building height feature, the hexagonal geometrical layout relationship, and the pairwise relationship of POIs. We show the importance of learning these spatial features in spatial data analysis applications, in particular, region similarity learning. We make the following contributions in this thesis: (1) We propose a CNN-based building height estimation method that computes building height from street scene images. We first detect roofline candidates from such images. Then, we use a deep neural network called RoofNet to filter these candidates and select the best candidate via an entropy-based ranking algorithm. When the true roofline is identified, we compute building height via the pinhole camera model. Experimental results show that our overall building height estimation method is more accurate than the baseline by up to 11.9%. (2) We propose a novel native hexagonal CNN framework named HexCNN for hexagonal layout relationship learning. HexCNN takes hexagon-shaped input and performs forward and backward propagation on the input based on hexagon-shaped filters, hence avoiding computation and memory overheads caused by transforming the input into rectangular shapes to fit traditional CNN models. Experimental results show that, compared with the state-of-the-art models, which imitate hexagonal processing but use rectangle-shaped filters, HexCNN reduces training time by up to 42.2%. Meanwhile, HexCNN saves memory by up to 25% and 41.7% for loading the input and performing convolution, respectively. (3) We propose a Graph Convolutional Network (GCN) model with weighted structural features named WGCN for learning edge direction features between POIs. WGCN first captures nodes' structural fingerprints via a direction and degree-aware Random Walk with Restart algorithm. The walk is guided by both edge direction and nodes' in- and out-degrees. Then, the interactions between nodes' structural fingerprints are used as the weighted node structural features. To further capture nodes' high-order dependencies and graph geometry, WGCN embeds graphs into a latent space to obtain nodes' latent neighbors and geometrical relationships. Based on nodes' geometrical relationships in the latent space, WGCN differentiates latent, in-neighbors, and out-neighbors with an attention-based geometrical aggregation. Experiments on transductive node classification tasks show that WGCN outperforms the baseline models consistently and by up to 17.07% in terms of accuracy on five benchmark datasets. (4) To showcase the effectiveness of learning the above spatial features in spatial data analysis applications, we propose a triplet-based Graph Neural Network model named C-MPGCN for region representation learning, where POIs are represented as graphs. C-MPGCN uses POIs' height and size as node features in addition to POIs' category and geo-location features used in existing methods. C-MPGCN also captures the POIs' hexagonal layout relationship and the pairwise relationship (edge features). Experimental results show that C-MPGCN outperforms the state-of-the-art methods for region similarity learning consistently under different evaluation matrices.
-
ItemMapping the structural connectome and predicting functional connectivity with deep learning methods( 2020)Mapping the human connectome is a major goal in neuroscience, where connectome refers to a comprehensive network description of the brain. This network is often represented as a graph, where nodes denote brain regions and edges represent white matter pathways. Tractography is a computational reconstruction method based on diffusion-weighted magnetic resonance imaging (dMRI) that estimates millions of streamlines that trace out the trajectories of white matter fiber bundles. The number of streamlines interconnecting each pair of regions comprising a predefined cortical parcellation is computed to yield a structural connectivity matrix. Network analyses of these connectivity matrices have yielded new insights into brain disorders (such as Schizophrenia, Alzheimer’s disease), cognition and neurodevelopmental processes. Moreover, the temporal dependence of neuronal activity patterns of different brain regions (functional connectivity) is also associated with underlying neuronal pathways (structural connectivity). In this thesis, we analyse the capabilities of state-of-the-art tractography algorithms (deterministic and probabilistic) for mapping connectomes and develop algorithms that overcome the limitations of conventional tractography algorithms for connectome mapping. Also, we utilize the structure-functional coupling for training Deep Neural Nets to predict the functional connectivity from structural connectivity. In the first part of the thesis, we develop numerical connectome phantoms that feature realistic network topologies and match to the fiber complexity of in vivo dMRI. The connectivity between pairs of regions was predefined for these phantoms. The phantoms are utilized to evaluate the performance of tensor-based and multi-fiber implementations of deterministic and probabilistic tractography. We found that multi-fiber deterministic tractography yields the most accurate connectome reconstructions, whereas probabilistic algorithms are hampered by an abundance of spurious connections. It is essential to omit connections with the fewest number of streamlines (thresholding) when using probabilistic algorithms for mapping connectomes. The study suggests that multi-fiber deterministic tractography is well suited for connectome mapping, regardless of the streamline threshold. In the second part, we propose a novel framework to map structural connectomes using deep learning. This framework not only enables connectome mapping with a convolutional neural network (CNN) but can also be straightforwardly incorporated into conventional connectome mapping pipelines (using tractography) to enhance accuracy. This framework involves decomposing the entire brain volume into overlapping blocks. Blocks are sufficiently small to ensure that a CNN can be efficiently trained to predict each block’s internal connectivity architecture. Later, a block stitching algorithm is proposed to rebuild the full brain volume from these blocks and thereby map end-to-end connectivity matrices. Performance is evaluated using simulated dMRI data generated from numerical connectome phantoms with known ground truth connectivity. Due to the redundancy achieved by allowing blocks to overlap, block decomposition and stitching steps can enhance the accuracy of probabilistic and deterministic tractography algorithms by up to 20-30%. Various studies have reported that functional brain connectivity is associated with underlying structural characteristics. In the third part of the thesis, we utilize this structure-functional coupling to develop a novel framework using deep learning that predicts functional connectivity from structural connectivity. The framework predicts functional connectivity without explicitly modelling the biophysical characteristics of the brain. We have demonstrated that a neural network can predict functional connectivity with high accuracy while preserving the inter-subject functional differences. Furthermore, we also demonstrated that functional connectivity could be used to predict human behavior, namely cognition. Altogether, the analyses and frameworks presented in this thesis aid in extracting structural connectivity and understanding the complex relationships between functional and structural connectivity in the human brain.