Browsing by Author "Yang, Hongzhou"
Now showing 1 - 7 of 7
Results Per Page
Sort Options
- ItemOpen AccessAgent-based Models in Flood Simulation - A Case Study for New Brunswick’s Flooding Events in 2018(2023-09-06) Zhu, Wenxuan; Stefanakis, Emmanuel; Yang, HongzhouThis thesis introduces an Agent-Based Modeling (ABM) framework for flood simulation and inversion modeling in flood-prone areas, aiming to improve our understanding of the complex dynamics of flooding and provide valuable insights for flood management. The developed model incorporates multi-source terrain datasets, and integrates water flow and meteorological conditions from remote sensing data sources. It takes into account factors such as precipitation patterns, geographical features, and Digital Elevation Model (DEM) resolution to accurately represent flood characteristics. The model's parameter settings are derived from extensive experimentation, allowing for effective control and meaningful results. By considering the impact of precipitation and the presence of rivers, the model demonstrates its ability to simulate flood inundation with a reasonable level of accuracy. Overall, this comprehensive model provides a valuable tool for flood simulation and offers insights into flood dynamics for effective flood management and mitigation strategies.
- ItemOpen AccessAutomatic Registration of Imagery to Mobile LiDAR Maps(2024-02-15) Jones, Kent Douglas; Lichti, Derek D.; Detchev, Ivan Denislavov; Yang, Hongzhou; El-Sheimy, Naser M.; Chapman, Michael AlastairMapping in 3D that records geospatial data from platform-mounted sensors with digital twinning supports maintenance and future planning of civil infrastructure. Three-dimensional mapping is efficiently performed with a Mobile Mapping System (MMS). This research demonstrates camera-only registration of subsequently captured images to an MMS point-cloud for updating MMS datasets. The research resolves key issues with inherent resolution differences between MMS laser scanner point-clouds and camera images by bridging differences between MMS point-clouds and camera images using a synthetic camera image (SCI). SCI are used to determine the approximate pose or coarse register the camera image to the MMS point cloud. The SCI coarse registration precision is maximized by generating surfaces, interpolating intensity values, and reducing noise with a median filter. The SCI is processed with a median filter to remove salt-and-pepper noise from the generation methods while preserving edges. Edgeboxes are adapted to find similar features in both SCIs and camera images. These features are then passed through layers of a convolutional neural network to provide a feature descriptor for coarse registration. Real camera images (RCI) are processed to mitigate resolution differences with the SCIs. The RCI is downsampled to align with the spatial resolution of the SCIs. Robust features are used to register the RCI to the SCIs. SIFT is used for fine registration between RCIs and SCIs generated from dense point-clouds. Landmark features are used for registration of RCIs to SCIs generated from MMS point-clouds. The edgebox parameters require tuning to detect the same features in two disparate image sets. The fourth layer of AlexNet was found to provide the most ideal feature descriptor for registration between RCIs and SCIs. The approximate location of the RCI using SCIs as interpreters between RCI and MMS point-cloud detect scenes at a precision of 97% when changes are less than 20%, and foliage does not exceed 20% of the camera image. This novel application of landmark features aligns with camera-to-camera place recognition precision. The focal length and IOPs do not influence the precision of the registration because the registration precision does not change when different cameras capture the real images.
- ItemOpen AccessDevelopment of a New Real-Time Precise Point Positioning System(2018-03-14) Yang, Hongzhou; Gao, Yang; El-Sheimy, Naser; Kattan, Lina; Skone, Susan H.; Bisnath, Sunil B.Real-time Precise Point Positioning (PPP) is drawing increasing attentions from both the Global Navigation Satellite System (GNSS) community and real-time users with different applications, such as offshore navigation, precise agriculture and hazard warning. To meet the rapidly increasing demand, the International GNSS Service (IGS) Real-Time Service (RTS) is currently disseminating several real-time high-frequency State Space Representation (SSR) products through the Internet under the Networked Transport of Radio Technical Commission for Maritime Services (RTCM) via the Internet Protocol (NTRIP) protocol. High availability of real-time PPP services requires high availability of precise orbit and clock corrections. Any correction outage, either due to corrupted ephemeris or loss of communication link, will degrade the availability of precise positioning using the service. Meanwhile, the communication burden is very heavy with such high update rates. To tackle the above limitations, a new robust real-time PPP system with higher availability is proposed in this thesis. The proposed system consists of three components regarding server end, communication end and user end. For the new real-time PPP system, the satellite orbit and clock Initial Parameters (IP) products are generated at the server end and broadcast to the user end for the generation of high precision orbit and clock products, afterwards, the real-time PPP can be carried out with the IP-based high precision satellite products. With the IP products, the real-time PPP system can operate with scalable update rates according to the various accuracy requests of different applications. Furthermore, the new real-time PPP system can continue during Internet connection outages, which is not uncommon in real applications due to the Internet connection losses. The prototype of the new real-time PPP system is developed and substantially tested in real-time.
- ItemOpen AccessMultiple Frame Point Cloud Object Detection Using Feature Fusion(2023-09-21) Huang, Minyuan; Leung, Henry; Yang, Hongzhou; Carriere, JayObject detection has important applications in the field of autonomous driving. While 2D object detection lacks depth information and suffers from the scale problems of objects, 3D point cloud object detection is a more suitable alternative for autonomous driving. However, due to the character of point cloud data, sparsity is an inevitable problem in point cloud object detection, and single-frame methods still have their limitations. This research investigates using multiple frame point cloud data to enhance detection precision. Facing the sparsity challenge in point cloud data, single-frame detection methods perform detection independently without all the data together to improve detection. However, with proper fusion scheme, using multiple frame data can be similar to using denser lidar. In this work, multiple association and fusion methods are discussed. And finally a probability-based proposal feature fusion (PPFF) module is proposed. Chapter 3 proposes a two-stage point cloud object detection method using proposal feature fusion. Proposal association methods and feature fusion methods are discussed. In Chapter 4, the MHT tracker is discussed, and in Chapter 5, combined with the knowledge of MHT, a probability-based proposal feature fusion (PPFF) module for multiple-frame point cloud object detection is proposed. We compared our methods to current multiple-frame methods to show the improvement and effectiveness of our methods.
- ItemOpen AccessNavigation Sensor Stochastic Error Modeling and Nonlinear Estimation for Low-Cost Land Vehicle Navigation(2023-09-12) Minaretzis, Chrysostomos; El-Sheimy, Naser; Noureldin, Aboelmagd; Gao, Yang; Yang, Hongzhou; Hefnawi, MostafaThe increasing use of low-cost inertial sensors in various mass-market applications necessitates their accurate stochastic modeling. Such task faces challenges due to outliers in the sensor measurements caused by internal and/or external factors. To optimize the navigation performance, robust estimation techniques are required to reduce the influence of outliers to the stochastic modeling process. The Generalized Method of Wavelet Moments (GMWM) and its Multi-signal extensions (MS-GMWM) represent the latest trend in the field of inertial sensor error stochastic analysis, they are capable of efficiently modeling the highly complex random errors displayed by low-cost and consumer-grade inertial sensors and provide very advantageous guarantees for the statistical properties of their estimation products. On the other hand, even though a robust version exists (RGMWM) for the single-signal method in order to protect the estimation process from the influence of outliers, their detection remains a challenging task, while such attribute has not yet been bestowed in the multi-signal approach. Moreover, the current implementation of the GMWM algorithm can be computationally intensive and does not provide the simplest (composite) model. In this work, a simplified implementation of the GMWM-based algorithm is presented along with techniques to reduce the complexity of the derived stochastic model under certain conditions. Also, it is shown via simulations that using the RGMWM every time, without the need for contamination existence confirmation, is a worthwhile trade-off between reducing the outlier effects and decreasing the estimator efficiency. Generally, stochastic modeling techniques, including the GMWM, make use of individual static signals for inference. However, it has been observed that when multiple static signal replicates are collected under the same conditions, they maintain the same model structure but exhibit variations in parameter values, a fact that called for the MS-GMWM. Here, a robust multi-signal method is introduced, based on the established GMWM framework and the Average Wavelet Variance (AWV) estimator, which encompasses two robustness levels: one for protection against outliers in each considered replicate and one to safeguard the estimation against the collection of signal replicates with significantly different behaviour than the majority. From that, two estimators are formulated, the Singly Robust AWV (SR-AWV) and the Doubly Robust (DR-AWV) and their model parameter estimation efficiency is confirmed under different data contamination scenarios in simulation and case studies. Furthermore, a hybrid case study is conducted that establishes a connection between model parameter estimation quality and implied navigation performance in those data contamination settings. Finally, the performance of the new technique is compared to the conventional Allan Variance in a land vehicle navigation experiment, where the inertial information is fused with an auxiliary source and vehicle movement constraints using the Extended and Unscented Kalman Filters (EKF/UKF). Notably, the results indicate that under linear-static conditions, the UKF with the new method provides a 16.8-17.3% improvement in 3D orientation compared to the conventional setting (AV with EKF), while the EKF gives a 7.5-9.7% improvement. Also, in dynamic conditions (i.e., turns), the UKF demonstrates an 14.7-17.8% improvement in horizontal positioning and an 11.9-12.5% in terms of 3D orientation, while the EKF has an 8.3-12.8% and an 11.4-11.7% improvement respectively. Overall, the UKF appears to perform better but has a significantly higher computational load compared to the EKF. Hence, the EKF appears to be a more realistic option for real-time applications such as autonomous vehicle navigation.
- ItemOpen AccessSelf-Supervised Learning Method for Semantic Segmentation of LiDAR Point Clouds(2024-04-23) Mutlu Kipirti, Fatma; Wang, Ruisheng; Wang, Ruisheng; Hassan, Quazi Khalid; Wang, Xin; Yang, HongzhouSemantic segmentation has shown a significant success for achieving comprehensive scene understanding in real-time perception and urban modeling. Over the recent years, there have been significant advancements in semantic segmentation for LiDAR point clouds, largely the adopting of deep learning techniques. There are the related works of 3D semantic segmentation, including neural network models to process converted voxels, points, and graphs. However, point-based methods are not taken into account local structure feature, resulting in a lack of fine-grained features and limited generalization. Additionally, these models do not take full advantage of the high-level geometric correlations among local neighbors, resulting in low semantic segmentation accuracy. The use of voxel-based methods for balancing precision and computational efficiency is a useful technique. Still, voxel-based representation of point clouds is inefficient and tends to ignore fine details. Little research has investigated using graph-based methods for LiDAR point cloud semantic segmentation. Our work demonstrates the feasibility of using graph representation for highly accurate semantic segmentation in a point cloud. Other problems in the existing methods are high computational and memory requirements. Self-supervised learning on large unlabeled datasets is one way to reduce the number of manual annotations needed. In this thesis, we explore that leverage the combination self-supervised contrastive learning and graph-based method to overcome the semantic segmentation challenges of large-scale point clouds. An experimental qualitative and quantitative analysis of our method shows that the proposed approach can beat previous approaches on S3DIS and SemanticKITTI datasets for the task of LiDAR point cloud semantic segmentation.
- ItemOpen AccessSequential Importance Resampling Particle Filter for Ambiguity Resolution(2023-08-30) Manzano Islas, Roberto Rene; O'Keefe, Kyle; El-Sheimy, Naser; Gao, Yang; Yang, Hongzhou; Bisnath, SunilIn this thesis the sequential importance resampling particle filter for estimating the full geometry-based float solution state vector for Global Navigation Satellite System (GNSS) ambiguity resolution is implemented. The full geometry-based state vector, consisting on position, velocity, acceleration, and float ambiguities, is estimated using a particle filter in RTK mode. In contrast to utilizing multi-frequency and multi-constellation GNSS measurements, this study employed solely L1 GPS code and carrier phase observations. This approach simulates scenarios wherein the signal reception environment is suboptimal and only a restricted number of satellites are visible. However, it should be noted that the methodology outlined in this thesis can be expanded for cases involving multiple frequencies and constellations. The distribution of particles after the resampling step is used to compute an empirical covariance matrix Pk based on the incorporated observations at each epoch. This covariance matrix is then used to transform the distribution using the decorrelating Z transformation of the LAMBDA method [1]. The performance of a float solution based on point mass representation is compared to the typically used extended Kalman filter (EKF) for searching the integer ambiguities using the three common search methods described in [2]: Integer Rounding, Integer Bootstrapping, and Integer Least Squares with and without the Z transformation. As Bayesian estimators are able to include highly non-linear elements and accurately describe non-Gaussian posterior densities, the particle filter outperforms the EKF when a constraint leading to highly non-Gaussian distributions is added to the estimator. Such is the case of the map-aiding constraint, which integrates digital road maps with GPS observations to compute a more accurate position state. The comparison between the position accuracy of the particle filter solution with and without the map-aiding constraint to the solution estimated with the EKF is made. The algorithm is tested in different segments of data and shows how the position convergence improves when adding digital road map information within the first thirty seconds of initializing the Particle Filter in different scenarios that include driving in a straight line, turning, and changing lanes. The assessment of the effect of the map-aiding algorithm on the ambiguity domain is carried out as well and it is shown how the convergence time of the float ambiguities improves when the position accuracy is improved by the constraint. The particle filter is able to weight the measurements according to any kind of distribution, unlike the EKF which always assumes a Gaussian distribution. The performance of the PF when having non-Gaussian measurements is assessed, such as when the measurements are distorted by multipath. Two additional steps are implemented, an outlier detection technique based on the predicted set of particles, and the use of a mixture of Gaussians to weight the measurements detected as outliers. The implemented outlier detection algorithm is based on the residual (or innovation) testing technique which is commonly applied into the EKF. The innovation and its covariance matrix are estimated from a predicted set of residuals using the transitional prior distribution and the measurement model. Then, the innovation is compared against the critical value of N (0, 1) at a level of significance α. The mixture of Gaussians is the weighted sum of two Gaussians, one from the measurement noise matrix, and the second being a scaled version of the first one describing the multipath error. This procedure de-weights the measurements with multipath, and reduces the bias in the position estimate. The proposed map-aiding algorithm improves the ambiguity convergence time by approximately 80%, while the deweighting process enhances it by around 25% for the segments of the vehicle dataset that were analyzed. This work serves as a demonstration of cases wherein the particle filter addresses the limitations of the EKF in estimating the float solution in ambiguity resolution. Such limitations include constraints that give rise to non-Gaussian probability density functions and the utilization of a distinct likelihood function for outlier measurements, as opposed to the Gaussian assumption made by the EKF. The proposed map-aided particle filter can be implemented in real-time to enhance the float ambiguity during the initial epochs after the filter has been initialized. This implementation proves beneficial in urban environments where there is a loss or complete obstruction of the GNSS signal.