David Lattanzi

This study presents a scene-based clustering methodology for underwater structural health monitoring (SHM), categorizing images by intrinsic scene attributes (visibility, color, and turbidity). Shifting focus from object detection to holistic scene understanding enhances automated condition assessment in underwater infrastructure inspection by prefiltering suboptimal images and optimizing downstream tasks such as image matching and alignment. Ensuring algorithms operate on images with consistent lighting, texture, and clarity reduces ambiguities and enhances feature correspondence. It integrates local and global feature extraction, correlation analysis, feature selection, dimensionality reduction, and outlier detection to improve clustering accuracy and pipeline robustness. Clustering uses hierarchical and centroid-based algorithms, evaluated by Silhouette Score, Calinski-Harabasz Index, F1-score, V-measure, Adjusted Mutual Information, and Adjusted Rand Index, applying the Hungarian algorithm for label mapping in pseudo-labeled cases. The methodology is validated on real and synthetic underwater datasets, including a bridge inspection video dataset with abovewater and underwater segments, and controlled laboratory experiments assessing turbidity effects on submerged structural elements. A key contribution is assessing feature matching performance across clustered and unclustered image sets, demonstrating that scene-based clustering significantly enhances keypoint detection, feature correspondence, and inlier ratios while reducing mismatches from illumination and viewpoint variations. Additionally, engineered features outperform pretrained deep learning features in this unsupervised setting, providing more reliable global scene representation and improved clustering and feature matching outcomes. These findings establish the proposed pipeline as a valuable tool for underwater SHM, supporting applications such as 3D reconstruction, infrastructure assessment, and autonomous underwater inspections.

Accurate and consistent forensic bruise assessment is critical in ensuring positive clinical and legal outcomes for victims of violence. In this study, a framework for automated bruise detection is presented that, for the first time, integrates narrowband alternate-light-source (ALS) forensic imaging and ambient white light imaging. This evaluation framework is designed to address long-standing issues with respect to equitable performance across skin tones and lighting scenarios via a combination of novel model diagnostic strategies. In particular, skin-tone balancing during training and testing, threshold-sensitivity analysis, and embedding-similarity partitioning are employed to quantify the model robustness and deployment trade-offs that arise in forensic image analysis. Models were implemented with ImageNet-pretrained backbones and trained on a unique, multi-annotator full-consensus dataset comprising both white-light and ALS (415 nm and 450 nm) images. The protocol emphasizes three axes of operational relevance: (1) illumination composition in training (W/ALS ratio); (2) subgroup fairness via targeted balancing; and (3) model operating-point selection (confidence and IoU thresholds) informed by confidence-stability metrics and bootstrapped uncertainty estimates. Systematic W/ALS ratio sweeps indicate peak accuracy under ALS-dominant training and declining performance as the proportion of white-light images increases within the training set. Skin-tone balancing reduced failure rates for darker skin tones but increased overprediction in some demographic subgroups. Embedding-similarity and seen/unseen injury analyses demonstrate inflated generalization under image-level partitioning. Ultimately, the findings suggest that future researchers and developers should employ injury-level data partitioning and ensure a weighted balance of ALS images during training.

Traffic speed forecasting is an important task in intelligent transportation system management. In current forecasting models, information modalities other than speed priors are largely not considered. While state of the art performance is achieved on speed forecasting with graph neural network methods, these methods do not incorporate information on roadway maintenance work zones and their impacts on predicted traffic flows; yet the impacts of construction work zones are of significant interest to roadway management agencies because they translate to impacts on the local economy and public well‐being. In this paper, we build over the convolutional graph neural network architecture and present a new ‘Graph Convolutional Network for Roadway Work Zones’ model that includes a novel data fusion mechanism and a new heterogeneous graph aggregation methodology to accommodate work zone information in spatio‐temporal dependencies among traffic states. The model is evaluated on two data sets that capture traffic flows in the presence of work zones in the Commonwealth of Virginia. Extensive comparative evaluation and ablation studies show that the proposed model can capture complex and nonlinear spatio‐temporal relationships across a transportation corridor, outperforming baseline models, particularly when predicting traffic flow during a work zone event.

Structural health monitoring (SHM) is one of the most important solutions to the ongoing problem of deteriorating infrastructure. A crucial type of data needed for SHM is the displacement response of structures under loading, as these can be directly related to integrity and system performance, or can be used indirectly in analysis of the infrastructure. Video-based displacement measurement has become an increasingly accessible and viable approach to SHM. There is now an array of computer vision methods that extract displacement measurements using different components of the underlying video signal. However, all of these approaches are known to suffer from high levels of noise relative to conventional sensor installations. This study considers how to employ deep learning-based data fusion that combines measurements from several video analysis approaches to enhance the signal to noise ratio (SNR), resulting in an ensemble measurement approach. The emphasis of the study is on a convolutional neural network (CNN) architecture. A generative adversarial network (GAN) approach was also evaluated but was not able to reliably converge during model training. The CNN approach was evaluated using an experimental dataset of vibrations from a lab-scaled structural system with varying dynamic properties. The results show that the CNN fusion approach improves the loss function significantly and serves to effectively denoise the individual signals extracted from the video.

Traffic speed forecasting is an important task in intelligent transportation system management. The objective of much of the current computational research is to minimize the difference between predicted and actual speeds, but information modalities other than speed priors are largely not taken into account. In particular, though state of the art performance is achieved on speed forecasting with graph neural network methods, these methods do not incorporate information on roadway maintenance work zones and their impacts on predicted traffic flows; yet, the impacts of construction work zones are of significant interest to roadway management agencies, because they translate to impacts on the local economy and public well-being. In this paper, we build over the convolutional graph neural network architecture and present a novel ``Graph Convolutional Network for Roadway Work Zones" model that includes a novel data fusion mechanism and a new heterogeneous graph aggregation methodology to accommodate work zone information in spatio-temporal dependencies among traffic states. The model is evaluated on two data sets that capture traffic flows in the presence of work zones in the Commonwealth of Virginia. Extensive comparative evaluation and ablation studies show that the proposed model can capture complex and nonlinear spatio-temporal relationships across a transportation corridor, outperforming baseline models, particularly when predicting traffic flow during a workzone event.

The ubiquity of smartphones has led to a variety of studies on how phone accelerometers can be used to assess road pavement quality. The majority of prior studies have emphasized supervised machine learning, assuming the ability to collect labeled data for model development. However, variances in vehicle dynamics and roadway quality, as well as the reliance on data labeling, limit the generalizability, scalability, and reproducibility of these approaches. Here, we propose an unsupervised learning framework that combines Pareto-optimized wavelet featurization and clustering. We first demonstrate the applicability of wavelets as features for dimensionally reduced accelerometer data. Wavelet featurization typically requires significant empirical tuning for optimal featurization. The presented framework automates tuning and selection of wavelet features based on subsequent clustering metrics. These metrics are related to inherent cluster characteristics such as intercluster variance and between-cluster variance. Rather than select a clustering method a priori, the proposed approach optimizes across a variety of clustering algorithms and hyperparameter configurations, again based on a set of clustering metrics. Experimental evaluation shows that the framework is able to detect road pavement distress and distinguish between classes of pavement defects, but that low-cost smartphone sensor data may not be as reliable in discriminating the more nuanced characteristics of pavement distress. This was most notable in the case of cracking, where the type and range of cracking severity caused undesirable cluster separation. The presented framework is general and cost-efficient and opens the way to further research on automatic pavement distress recognition from crowdsourced, low-cost data.

Recent improvements in remote sensing technologies have shown that techniques such as photogrammetry and laser scanning can resolve geometric details at the millimeter scale. This is significant because it has expanded the range of structural health monitoring scenarios where these techniques can be used. In this work, we explore how 3D geometric measurements extracted from photogrammetric point clouds can be used to evaluate the performance of a highway bridge during a static load test. Various point cloud registration and deformation tracking algorithms are explored. Included is an introduction to a novel deformation tracking algorithm that uses the interpolation technique of kriging as the basis for measuring the geometric changes. The challenging nature of 3D point cloud data means that statistical methods must be employed to adequately evaluate the deformation field of the bridge. The results demonstrate a pathway from the collection of digital photographs to a mechanical analysis with results that capture the bridge deformation within one standard deviation of the mean reported value. These results are promising given that the midspan bridge deformation for the load test is only a few millimeters. Ultimately, the approaches evaluated in this work yielded errors on the order of 1 mm or less for ground truth deflections as small as 3.5 mm. Future work for this method will investigate using these results for updating finite element models.

For automated robotic manufacturing, a key aspect of monitoring is the identification and segmentation of core actuation processes captured in sensor logs. Once segmented, the behavior of an industrial system during a particular actuation can be tracked to detect signs of degradation. This study presents a technique for performing such an analysis through a combination of machine learning techniques designed to work with an acoustic monitoring system. A spectrogram-based convolutional neural network (CNN) is first trained to identify and segment primary motion classes from acoustic data. Unsupervised clustering and feature-space analysis are then employed to further separate the data into motion sub-classes beyond the capabilities of the CNN. This approach was evaluated on acoustic recordings of a Selective Compliance Assembly Robot Arm (SCARA) system. The developed CNN performed primary robotic motion segmentation with a maximum actuation identification accuracy of 87% when compared to validation data. The unsupervised clustering process had mixed success in distinguishing more fine-grained motion sub-classes due to strong variances in signal energy for some sub-classes. Further refinement is required for improved segmentation accuracy as well as automatic feature generation. The application of this process for life-cycle system monitoring is discussed as well.