Urbanization over the next decade will present many complex challenges to cities. Transportation plays an important role in a city’s ability to address urbanization and create resilient, sustainable, inclusive, and safer cities for all. However, challenges introduced by urbanization will decrease the effectiveness of current transportation methodologies. By utilizing innovations in sensing, computing, and communication along with machine learning and real-time data, intelligent transportation systems (ITS) can advance transportation efficiency and effectiveness. Integrating these technologies into intelligent transportation systems can potentially reduce congestion, pollution, energy consumption, and traffic incidents. This integration is at the forefront of academic and industry research. Although testbeds have been developed for the development, testing, and validation of ITS solutions, existing testbeds focus on narrow aspects of the ITS domain and are unable to support large-scale experimentation and integrations. Additionally, access to resources and data is limited and not remotely accessible to potential researchers. In this paper, we propose a smart city and ITS Testbed-As-A-Service. The testbed is located in downtown Chattanooga, TN, and offers a unified platform for a wide range of ITS applications including connected and autonomous vehicles, wireless communication, cooperative transportation, internet-of-things (IoT), and edge-based AI.

Electric vehicle (EV) charging stations are critical in promoting EV adoption and mitigating global warming by reducing reliance on fossil fuels. However, a comprehensive understanding of the latent characteristics of these charging stations remains limited. Unveiling these latent features is essential for enhancing predictive tasks such as utilization prediction, demand forecasting, and strategic infrastructure planning. In this paper, we conduct a comprehensive investigation into methods for extracting embedding vectors of charging stations based on userstation interactions. We explore a spectrum of techniques—from traditional approaches like non-negative matrix factorization to advanced machine learning models such as neural collaborative filtering—to effectively capture these latent features. Through extensive experiments and analyses, we evaluate the quality and effectiveness of the generated embeddings in improving predictive modeling tasks related to charging station usage. Our findings demonstrate that incorporating these embeddings significantly enhances the performance of predictive models, leading to more accurate demand forecasts and better utilization predictions. To the best of our knowledge, this is the first study to delve deeply into extracting and analyzing embedding vectors of charging stations derived from user interaction data. The insights gained from this research provide valuable guidance for optimizing EV charging infrastructure and can inform future developments in the field, ultimately supporting the broader adoption of electric vehicles.

Accurate speed estimation of vehicles is crucial for applications like congestion management and speed limit enforcement to reduce accidents. Despite its relevance, datasets for visual speed estimation are limited. In this study, we propose a deep learning-based framework for vehicle tracking and visual speed estimation, implemented on the MLK Smart Corridor. Pretrained YOLOV7 was used for vehicles detection. Furthermore, we enhanced the DeepSort tracking model by replacing its feature extractor with a Siamese network trained on the UADETRAC dataset. This modification resulted in a 71% decrease in the trajectories IDSW rate when evaluated against the KITTI benchmark. Our framework can also estimate vehicle speeds using reference object scaling and perspective transformation, achieving a commendable error rate of 0.516 mph. Recognizing the potential of our speed estimator, we identified the need to build a new dataset for visual speed estimation. For this task, a vehicle equipped with an OBU is used to collect speed ground-truth data, and videos were captured simultaneously and passed to the framework for estimating speed readings to be validated. The final dataset contains a total of 1326 annotated instances stamped with their speed readings and grouped into 78 trajectories. This dataset is intended to be used as a benchmark to facilitate research in speed estimation from visual data obtained from urban testbeds.

The public commute is essential to all urban centers and is an efficient and environment-friendly way to travel. Transit systems must become more accessible and user-friendly. Since public transit is majorly designed statically, with very few improvements coming over time, it can get stagnated, unable to update itself with changing population trends. To better understand transportation demands and make them more usable, efficient, and demographic-focused, we propose a fast, multi-layered transit simulation that primarily focuses on public transit simulation (BTE-Sim). BTE-Sim is designed based on the population demand, existing traffic conditions, and the road networks that exist in a region. The system is versatile, with the ability to run different configurations of the existing transit routes, or inculcate any new changes that may seem necessary, or even in extreme cases, new transit network design as well. In all situations, it can compare multiple transit networks and provide evaluation metrics for them. It provides detailed data on each transit vehicle, the trips it performs, its on-time performance and other necessary factors. Its highlighting feature is the considerably low computation time it requires to perform all these tasks and provide consistently reliable results.

Traffic simulators have many parameters that describe vehicle characteristics and driving behaviors. However, driving behaviors differ across urban, suburban, and rural areas. Even in the same area, driving behavior can be affected by the time of day or weather conditions. Therefore, it is difficult to get an accurate parameter set that is suitable for all scenarios. As a result, default parameters of simulators are usually determined only for a specific test case. To simulate a traffic scenario, researchers need to perform calibration to determine a suitable parameter set, which can provide more reliable simulated traffic than the default parameter set. A popular approach is manual calibration using human experience, but it is usually not effective due to the huge space of possible parameters. Although some studies proposed automated methods using evolutionary algorithms, implementing these methods is a time-consuming job. In this paper, we introduce a toolkit for researchers to easily conduct calibration for their own traffic scenarios. The toolkit supports several state-of-the-art algorithms and is designed to run in parallel to utilize the power of high performance computers. Moreover, by using this toolkit, we conduct in-depth experiments to understand which factors affect the calibration performance.

The volume and velocity of traffic data have in-creased dramatically due to the wide adoption of new technologies such as cameras, Internet-of-Thing devices, and vehicular net-works. That data can help us to optimize Traffic Signal Controls (TSCs) by using adaptive algorithms. Some direct applications of these algorithms are reducing the CO2 emission, fuel consumption, and traveling time. Recently, Deep Reinforcement Learning (DRL) methods are the de-facto solution due to its ability to handle big data with high performance. However, most open source codes and frameworks for TSCs using DRL algorithms have limited flexibility. That causes a difficulty to reuse the codebases for new contexts. Therefore, it will be difficult to have a benchmark for TSCs using different optimization algorithms. For this reason, our paper introduces TSLib – a Python framework for fast prototyping TSCs. Specifically, TSLib is designed as a modular system with high reusability so that researchers can quickly implement and evaluate new ideas of TSCs. Moreover, our work offers a comprehensive implementation of some well known TSCs algorithm including both traditional and DRL-based methods as well as their performance measurements.

The COVID-19 pandemic has caused a drastic change in traffic in the U.S and throughout the world. With the drop in traffic volume, the fuel consumption used for traveling should decline. This study focuses on the changes in fuel consumption of public transportation before and during the pandemic. The fuel consumption volumes of diesel bus fleet in Chattanooga were analyzed to identify the changes between the two periods. Our study provides preparation for future disasters.

Stroke is the fifth leading cause of death for Americans. Due to critical consequences and cost, an efficient stroke system of care (i.e., acute treatment, post-stroke acute rehabilitation) is necessary. Early determination of hospital discharge disposition is important for stroke management, and can make an immense impact in optimizing acute treatment and planning post-acute rehabilitation with desired outcomes. With the rise of sophisticated machine learning models, many researchers have gained momentum in using such models in their studies. However, due to the lack of explanations that models provide, the integrity of the prediction result is often challenged. Our goal is to predict post-stroke hospital discharge disposition using machine learning models to increase the prediction capability while providing explanations for the results using an interpretation method, such as Local Interpretable Model-agnoistic Explanations (LIME). Our results demonstrate the effectiveness of LIME in providing interpretability to machine learning models and suggests further exploration in performance improvement.

Urbanization over the next decade will present many complex challenges to developing cities. The smart city concept aims to address these challenges by exploiting large scale deployments of Internet of Things (IoT) and communication technologies. These technologies generate data that provide quantifiable insights into the state of the infrastructure within a city. Using these insights, cities can more effectively allocate resources, manage services, and enhance the lives of its citizens. The data generated by smart cities is complex and requires high throughput. Advanced data integration platforms must support city-wide data collection, analysis, and storage. These systems must provide features that allow them to scale alongside the growth of the cities to support high rates of data ingestion in large volumes. Additionally, these systems must support low latency response times which is a critical requirement for time sensitive smart city applications. In this paper, we introduce a smart city testbed that will provide a real-world testing environment for applications in areas such as intelligent transportation, pedestrian safety, and autonomous vehicles. The proposed testbed will act as an open platform for researchers and developers to test new sensors, algorithms and more in a live urban environment, allowing them to test before deploying a product or application. In addition to the physical testbed and its capabilities, we will discuss the data integration system and applications responsible for collecting, analyzing, and storing the data generated by the testbed. Lastly, we will introduce an open data platform where researchers can access datasets generated by the testbed.

In smart cities equipped with cameras, one desirable use-case is to detect and track objects. While object detection has been implemented using various methods, object tracking poses a different problem; to track an object requires object permanence to be established between each frame of video. While many technologies have been proposed as a solution for problem, an implementation with scalability in mind has not been developed and poses many new challenges. This paper proposes e-SORT, a solution for scalable object tracking using an enhanced version of the Simple Online and Realtime Tracking (SORT) algorithm. Beyond its scalability, e-SORT stores a mapping of each objects' locations such that the full path of each object is available and several metrics (such as velocity and acceleration) can be calculated. Both e-SORT's abilities and our proposed solution to scalable object tracking are tested and evaluated on Chattanooga Tennessee's live urban testbed.

Monitoring energy consumption data is essential to the everyday workings of power companies; a single uncaught incident outside the standards of normal use can result in financial loss. To minimize the repercussions of an uncaught error, the utilization of forecasting and machine learning can significantly improve the detection of such anomalies in day-to-day operations. This study covers power anomaly detection with the use of deep learning algorithms that have the capability of removing seasonality and trend from data, yielding residual values that are applied in a comparison to values generated from predictive analysis using recurrent neural networks (RNN). Data for this study is provided by Tennessee Valley Authority (TVA).

To create safer and less congested traffic operating environments researchers at the University of Tennessee at Chattanooga (UTC) and the Georgia Tech Research Institute (GTRI) have fostered a vision of cooperative sensing and cooperative mobility. This vision is realized in a mobile application that combines visual data extracted from cameras on roadway infrastructure with a user's coordinates via a GPS-enabled device to create a visual representation of the driving or walking environment surrounding the application user. By merging the concepts of computer vision, object detection, and mono-vision image depth calculation, this application is able to gather absolute Global Positioning System (GPS) coordinates from a user's mobile device and combine them with relative GPS coordinates determined by the infrastructure cameras and determine the position of vehicles and pedestrians without the knowledge of their absolute GPS coordinates. The joined data is then used by an iOS mobile application to display a map showing the location of other entities such as vehicles, pedestrians, and obstacles creating a real-time visual representation of the surrounding area prior to the area appearing in the user's visual perspective. Furthermore, a feature was implemented to display routing by using the results of a traffic scenario that was analyzed by rerouting algorithms in a simulated environment. By displaying where proximal entities are concentrated and showing recommended optional routes, users have the ability to be more informed and aware when making traffic decisions helping ensure a higher level of overall safety on our roadways. This vision would not be possible without high speed gigabit network infrastructure installed in Chattanooga, Tennessee and UTC's wireless testbed, which was used to test many functions of this application. This network was required to reduce the latency of the massive amount of data generated by the infrastructure and vehicles that utilize the testbed; having results from this data come back in real-time is a critical component.

Falls are common and dangerous for the elderly or individuals with decreased independence or functional limitations. Fall recognition is extremely important for fallers, healthcare providers, and society. Immediate fall recognition triggers emergency services and potentially decreases individuals time with injury without care. Acute post-fall intervention works to mitigate life threatening fall consequences, decrease fall risk through rehabilitation, and improve quality of life. Extended from our research on real-time fall risk estimation with the functional reach test and Timed Up and Go test built in mStroke, a real-time and automatic mobile health system for post-stroke recovery and rehabilitation, our investigation here is expanded to include fall recognition by taking advantage of wearable technologies and machine learning algorithms. Up to three wearable sensors are employed to acquire raw motion data related to activities of daily living or falls. Feature selection and classification on the basis of machine learning algorithms are explored for fall recognition. The fall recognition performances are presented to justify their accuracy and reliability. Meanwhile, the effects of sensor placement/location and the feature number on the recognition performance are also discussed in this paper.

In this paper, we develop a distributed compression technique that has low decoding and encoding computational complexity. The proposed scheme exploits both temporal and spatial correlations between nodes in distributed sensor networks. In case of events occurring, the values of both spatial and temporal might change and the compression technique needs to adjust its rate to the changes automatically. Our proposed algorithm reactively changes its compression rate to adapt to the variations in the correlations. This algorithm uses the well-known compressive sensing algorithm to exploit the spatial correlation. Rate less codes were adopted to generate the measurements. The number of measurements are adjusted based on the temporal correlations among sensors. When sensor readings are changing slowly, the compression rate is improved by reducing the number of measurements. In case of any event that significantly changes the signal readings, the algorithm generates more measurements to guarantee recovery of signal at the base station. The experimental results done over data gathered by 64 temperature sensors and also Matlab simulation results reveal that our algorithm is flexible to adapt the variations in the sensor readings, while it keeps the compression rate the minimum.

In this paper, we study the problem of data acquisition in wireless sensor networks (WSNs). A recently revitalized technique called compressive sensing (CS) has presented a new method to capture sparse signals at a rate below Nyquist. There are drawbacks to directly applying the existing CS algorithm to WSNs, which are mainly due to the fact that CS requires a large number of inter-communications for generating each projection. To mitigate these drawbacks, we propose compressive distributed sensing using random walk (CDS(RW)), an algorithm for CS in WSNs that uses rate less coding. This algorithm is independent of routing algorithms and network topologies. CDS(RW) collects sufficient number of sensor readings while combining them together without significantly increasing the inter-communication cost. We model the CS problem with code design for a set of parallel channels which helps us to design the rate less code degree distribution. This model provides the advantage of using non-uniform and unequal error protection codes.

Inspired by our recent work on lossy distributed source coding with side information available at the decoder, we propose a practical scheme for information embedding system with side information available at the encoder. Our proposed scheme is based on sending parity bits using LDPC codes. We provide a design procedure for the LDPC code that guarantees performance close to the Gelfand-Pinsker and Wyner-Ziv limits. Using simulation results, we show that the proposed method performs close to both Wyner-Ziv and Gelfand-Pinsker theoretical limits for even short length codes.

Summary form only given. First, we consider a system of two statistically dependent signals, X/sub 1/ and X/sub 2/. The correlation between signals X/sub 1/ and X/sub 2/ is modeled as the input and output of a binary symmetric channel with crossover probability. We propose a non-uniform LDPC code that considers the fact that different bits are subject to different sources of noise. Simulation results show that our proposed method improves the source coding performance considerably. We further study an extension of our approach to three correlated nodes. For simplicity, we consider the case that assumes that the sources are pairwise correlated with the same correlation probability. We show that this assumption cannot be generalized to more than three sources. In this case, a fourth source is not be a random variable and it can be identified by the first three sources.