This paper presents DENSER, a framework leveraging 3D Gaussian splatting (3DGS) for the reconstruction of dynamic urban environments. While several methods for photorealistic scene representations, both implicitly using neural radiance fields (NeRF) and explicitly using 3DGS have shown promising results in scene reconstruction of relatively complex dynamic scenes, modeling the dynamic appearance of foreground objects tends to be challenging, limiting the applicability of these methods to capture subtleties and details of the scenes, especially for dynamic objects. To this end, we propose DENSER, a framework that significantly enhances the representation of dynamic objects and accurately models the appearance of dynamic objects in the driving scene. Instead of directly using Spherical Harmonics (SH) to model the appearance of dynamic objects, we introduce and integrate a new method aiming at dynamically estimating SH bases using wavelets, resulting in better representation of dynamic objects appearance in both space and time. Besides object appearance, DENSER enhances object shape representation through densification of its point cloud across multiple scene frames, resulting in faster convergence of model training. Extensive evaluations on the KITTI dataset show that the proposed approach outperforms state-of-the-art methods by a wide margin. Source codes and models will be uploaded to this repository https://github.com/sntubix/denser

Automated defect detection in photovoltaic (PV) modules is essential for their maintenance and efficiency, yet challenges such as limited and imbalanced datasets hinder the adoption of high-accuracy systems. This study evaluates six semantic segmentation architectures based on U-Net and SegNet, paired with VGG16, MobileNet, and ResNet50 encoders, and trained on the 29-class dataset of PV module electroluminescence (EL) images. To address dataset imbalance, custom class weights were applied for all the feature and defect classes. VGG16-UNet outperformed other architectures, achieving a mean intersection over union (IoU) of 0.663 for feature classes and 0.326 across defect classes. In particular, it improved the detection of rare defects, such as dead cell, by 0.129 IoU. While previous research focused on a specific subset of classes, this study is the first to provide a comprehensive performance evaluation across all classes. It establishes a baseline for multi-class semantic segmentation in PV defect detection, laying the groundwork for further industrial applications such as in-field defect detection integrated into solar panel cleaning robots. Our implementation is publicly available at https://github.com/sntubix/pv-defect-segmentation , facilitating further research and development.

High-Definition (HD) maps play a crucial role in autonomous vehicle navigation, complementing onboard perception sensors for improved accuracy and safety. Traditional HD map generation relies on dedicated mapping vehicles, which are costly and fail to capture real-time infrastructure changes. This paper presents HDMapLaneNet, a novel framework leveraging V2X communication and Scene Graph Generation to collaboratively construct a localized geometric layer of HD maps. The approach extracts lane centerlines from front-facing camera images, represents them as graphs, and transmits the data for global aggregation to the cloud via V2X. Preliminary results on the nuScenes dataset demonstrate superior association prediction performance compared to STSU as a state-of-the-art baseline method.

The ever increasing regulatory requirements for Anti-Money Laundering (AML) compliance presents significant challenges for financial institutions and small businesses world-wide. Efficiently navigating these requirements is crucial not only for legal adherence but also for safeguarding the integrity of the global financial system. In response to this challenge, we develop a framework that uses advanced algorithms to improve identification and risk assessment processes within Know Your Customer (KYC) procedures. Using a technique to measure graph-based node similarities, our approach enhances the detection of Politically Exposed Persons (PEPs) and their known associates, facilitating a more nuanced and comprehensive analysis than traditional methods allow. We study the dataset of Ultimate Beneficial Owner (UBO) registry in Luxembourg and translate our findings into two risk indicators: involvement with underage shareholders, and number of companies at the address. We integrate these two indicators, as well as several other components of AML compliance, including country risk indices, beneficial ownership structures, and adverse media exposure, into a single coherent risk metric. The framework is designed to be both modular, supporting various degrees of regulatory scrutiny, and scalable, suitable for evolving regulatory landscapes. This risk metric can be used to determine whether Enhanced Due Diligence (EDD) is required by European AML directives. The end result is a more robust defense against financial crimes and improved AML processes within the EU and beyond.

The development and testing of real-time perception capabilities is crucial to fully realize the potential of connected and automated driving. This study presents a comprehensive analysis of end-to-end delay and object detection quality for distributed perception. We investigate object detection from camera images in two distinct scenarios: local and cloud processing. In the local scenario, objects are detected using only on-board hardware. The results are then used to generate Cooperative Perception Messages (CPM), which are broadcast to nearby vehicles via the ITS-G5 communication technology. In the cloud processing scenario, images are compressed using H.265 before being transmitted to the cloud via C-V2X, where objects are detected on more demanding hardware. The detection results are then transmitted to vehicles in the vicinity, enabling cooperative perception scenarios. This study evaluates real-time object detection systems in real-world conditions, and highlights the trade-offs between the end-to-end detection delay and quality. By leveraging emerging technologies such as ITS-G5 and C-V2X, our research provides crucial insights for the development of efficient perception systems in the connected driving ecosystem.

Environmental perception is a key element of autonomous driving because the information received from the perception module influences core driving decisions. An outstanding challenge in real-time perception for autonomous driving lies in finding the best trade-off between detection quality and latency. Major constraints on both computation and power must be taken into account for real-time perception in autonomous vehicles. Larger detection models tend to produce the best results but are also slower at runtime. Since the most accurate detectors may not run in real-time locally, we investigate the possibility of offloading computation to edge and cloud platforms, which are less resource-constrained. We create a synthetic dataset to train object detection models and evaluate different offloading strategies. We measure inference and processing times for object detection on real hardware, and we rely on a network simulation framework to estimate data transmission latency. Our study compares different trade-offs between prediction quality and end-to-end delay. Following the existing literature, we aim to perform object detection at a rate of 20Hz. Since sending raw frames over the network implies additional transmission delays, we also explore the use of JPEG and H.265 compression at varying qualities and measure their impact on prediction. We show that models with adequate compression can be run in real-time on the edge/cloud while outperforming local detection performance.

Automated vehicles use sensors to perceive the environment, and studies have shown the limitations of these sensors. The onboard sensors may not detect objects when other participants occlude the Field of View (FoV). Thus, sensor efficiency must be tested to ensure its reliability. Simulation is an excellent test option due to the complexity associated with practical experiments. However, emerging simulation frameworks still have various limitations, especially when it comes to large-scale evaluation. This work investigates realistic perception simulation options for autonomous vehicles. We report the perception accuracy for different traffic scenarios.

Convolutional Neural Networks have been successfully used to steer vehicles using only road-facing cameras. In this work, we investigate the use of Privileged Information for training an end-to-end lane following model. Starting from the prior assumption that such a model should spend a sizeable fraction of its focus on lane markings, we take advantage of lane geometry information available at training time to improve its performance. To this end, we constrain the class of learnable functions by imposing a prior stemming from a lane segmentation task. For each input frame, we compute the set of pixels that most contribute to the prediction of the model using the VisualBackProp method. These pixel-relevance heatmaps are then compared with ground truth lane segmentation masks. A Distraction Loss term is added to the objective to regularize the training process. We learn from real-world data collected using our experimental vehicle and compare the results to those obtained using the simple Mean Squared Error objective. We show that the presence of our regularizer benefits both the performance and the stability of the model across a variety of evaluation metrics. We use a pretrained lane segmentation model without fine-tuning to extract lane marking masks and show that valuable learning signals can be extracted even from imperfect privileged knowledge. This method can be implemented easily and very efficiently on top of an existing architecture, without requiring the addition of any trainable parameter.

Robot Operating System (ROS) is becoming more and more important and is used widely by developers and researchers in various domains. One of the most important fields where it is being used is the self-driving cars industry. However, this framework is far from being totally secure, and the existing security breaches do not have robust solutions. In this paper we focus on the camera vulnerabilities, as it is often the most important source for the environment discovery and the decision-making process. We propose an unsupervised anomaly detection tool for detecting suspicious frames incoming from camera flows. Our solution is based on spatio-temporal autoencoders used to truthfully reconstruct the camera frames and detect abnormal ones by measuring the difference with the input. We test our approach on a real-word dataset, i.e. flows coming from embedded cameras of self-driving cars. Our solution outperforms the existing works on different scenarios.

We propose a novel way of linking mobile network signaling data to the state of the underlying urban road network. We show how a predictive model of traffic flows can be created from mobile network signaling data. To achieve this, we estimate the vehicular density inside specific areas using a polynomial function of the inner and exiting mobile phone handovers performed by the base stations covering those areas. We can then use the aggregated handovers as flow proxies alongside the density proxy to directly estimate an average velocity within an area. We evaluate the model in a simulation study of Luxembourg city and generalize our findings using a real-world data set extracted from the LTE network of a Luxembourg operator. By predicting the real traffic states as measured through floating car data, we achieve a mean absolute percentage error of 11.12%. Furthermore, in our study case, the approximations of the network macroscopic fundamental diagrams (MFD) of road network partitions can be generated. The analyzed data exhibit low variance with respect to a quadratic concave flow-density function, which is inline with the previous theoretical results on MFDs and are similar when estimated from simulation and real data. These results indicate that mobile signaling data can potentially be used to approximate MFDs of the underlying road network and contribute to better estimate road traffic states in urban congested networks.

The proliferation of connected mobile devices together with advances in their sensing capacity has enabled a new distributed telematics platform. In particular, smartphones can be used as driving sensors to identify individual driver behavior and risky maneuvers. However, in order to estimate driver behavior with smartphones, the system must deal with different vehicle characteristics. This is the main limitation of existing sensing platforms, which are principally based on fixed thresholds for different sensing parameters. In this paper, we propose an adaptive driving maneuver detection mechanism that iteratively builds a statistical model of the driver, vehicle, and smartphone combination using a multivariate normal model. By means of experimentation over a test track and public roads, we first explore the capacity of different sensor input combinations to detect risky driving maneuvers, and we propose a training mechanism that adapts the profiling model to the vehicle, driver, and road topology. A large-scale evaluation study is conducted, showing that the model for maneuver detection and scoring is able to adapt to different drivers, vehicles, and road conditions.

Recent technological advances have allowed the development of miniaturized sensors and the emergence of a wide range of connected objects. Whether it's smartphones or in the broader sense wearables, the diversity of these devices and their accessibility opens up new fields for applications in the computer sciences [2, 3]. Smartwatches, which are experiencing a boom on the market, will be integral to the research that will shape the Internet in the years to come, namely big data, sensing systems and human behavior. Our demonstration falls within this context and aims to demonstrate the potential of these emerging technologies to respond to problems and to way of thinking introduced by industry and the scientific community, which are generally limited to smartphone sensing frameworks [1]. Further, we plan to present our research platform, SWIPE, which is dedicated to collecting, studying and learning about human dynamics by means of an ecosystem of wearables. A short presentation video is available online at http://swipe.sfaye.com/mobisys15/.

We address the problem of shock wave formation in uncoordinated highway traffic. First, we identify the combination of heavy traffic and small traffic perturbations or unexpected driver actions as the main causes of highway traffic jams. Then we introduce a novel distributed communication protocol that enables us to eliminate upstream shock wave formation even with low system penetration rates. Based on traffic information ahead, we propose a Cooperative Advanced Driver Assistance System (CADAS) that recommends non-intuitive velocity reductions in order to redistribute traffic more uniformly thereby eliminating traffic peaks. Simulation results show that CADAS significantly increases the average velocity and therewith reduces the overall travel time and avoids unnecessary slowdowns.

Smartphones embed increasingly complex sensors that can be used for a wide range of connected applications. Their growing market penetration provides the opportunity to develop novel distributed sensing platforms that will constitute the foundation for emerging commercial applications. In this work we present SenseFleet, a smartphone-based driver profiling platform. The concept behind SenseFleet is to analyze the output of smartphone sensors while driving in order to identify risky maneuvers (e.g. acceleration, braking and cornering) and provide a score reflecting the overall performance of the driver. A Fuzzy Inference System (FIS) is used to detect events independently from the car and mobile phone used. A smartphone application has been developed to demonstrate the feasibility of our approach. In this demo we show interactive results of a test campaign performed in Luxembourg. The results show that the SenseFleet platform is able to identify risky driving maneuvers and provide representative scores. A second benefit of SenseFleet is that for a large number of participants, it will allow to monitor the road network and identify dangerous locations, i.e. locations that have a high event density.

Profiling driving behavior has become a relevant aspect in fleet management, automotive insurance and eco-driving. Detecting inefficient or aggressive drivers can help reducing fleet degradation, insurance policy cost and fuel consumption. In this paper, we present a Fuzzy-Logic based driver scoring mechanism that uses smartphone sensing data, including accelerometers and GPS. In order to evaluate the proposed mechanism, we have collected traces from a testbed consisting in 20 vehicles equipped with an Android sensing application we have developed to this end. The results show that the proposed sensing variables using smartphones can be merged to provide each driver with a single score.

Over the last few years vehicular traffic density has continuously increased and it is likely that the traffic demand will further increase in the future. In many metropolitan areas the road infrastructure is no longer able to handle the peak traffic demand and the existing road network cannot be easily extended. The best way to improve the traffic situation is to optimise the resources available in the transportation network and to coordinate the traffic demand. The approach that we propose in this paper makes use of a collaborative traffic coordination protocol which collects real-time Floating Car Data (FCD) directly from participating vehicles and suggests dynamic routes in order to minimize travel delay. Information such as speed, position and direction is sent to one or more Traffic Coordination Points (TCPs) where it is aggregated to obtain a global picture of the traffic conditions in real-time. Based on this data, we continuously compute the route that minimizes the travel time to a given destination by applying Wardrop's first principle of equilibrium. Our results show that, by coordinating the vehicles, we are able to better distribute the overall traffic demand throughout the transportation network, reducing the average travel times and accommodating more vehicles.

The following paper provides a complete overview of LuxTraffic, an online system for collaborative traffic sensing in Luxembourg. LuxTraffic has been created to gather, analyze and archive anonymous user generated traffic data by taking advantage of the communication capabilities and the built-in GPS receiver of smartphone devices. We motivate the necessity and benefits of such a system, reveal the core functionalities as well as the technical details behind its realization. Next, we continue with a statistical evaluation about the system usage. We conclude by discussing the the pros and cons of such a system and propose future works.

In this paper we address the problem of uncoordinated highway traffic. We first identify the main causes of the capacity drop, namely high traffic demand and inadequate driver reaction. In the past, traffic and user behavior have been accurately described by cellular automata (CA) models. In this paper we extend the CA model to deal with highway traffic fluctuations and jams. Specifically, the model incorporates the communication layer between vehicles. The model thus enables us to study the impact of inter-vehicular communications and in particular the delivery of critical and timely upstream traffic information on driver reaction. Based on the newly-available traffic metrics, we propose an Advanced Driver Assistance System (ADAS) that suggests non-intuitive speed reduction in order to avoid the formation of so-called phantom jams. The results show that using such a system considerably increases the overall traffic flow, reduces travel time and avoids unnecessary slow-downs.

Advances in portable technologies and emergence of new applications stimulate interest in urban vehicular communications for commercial, military, and homeland defense applications. Simulation is an essential tool to study the behavior and evaluate the performance of protocols and applications in large-scale urban vehicular ad hoc networks (VANET). In this paper, we propose CORNER, a low computational cost yet accurate urban propagation model for mobile networks. CORNER estimates the presence of buildings and obstacles along the signal path using information extrapolated from urban digital maps. A reverse geocoding algorithm is used to classify the propagation situation of any two nodes that need to communicate starting from their geographical coordinates. We classify the relative position of the sender and the receiver as in line of sight (LOS) or nonline of sight (NLOS). Based on this classification, we apply different formulas to compute the path loss (PL) metric. CORNER has been validated through extensive on-the-road experiments, the results show high accuracy in predicting the network connectivity. In addition, on-the-road experiments suggest the need to refine the fading model to differentiate between LOS, and NLOS situations. Finally, we show the impact of CORNER on simulation results for widely used applications.

In the near future vehicular networks based on wireless technology will be part of our lives. Efficient and robust routing algorithms will play a key role in the success of such technology. In this paper we present TrafRoute, an efficient and robust routing scheme for vehicular networks, suitable for both Vehicle-to-Vehicle and Vehicle-to-Infrastructure communications. TrafRoute introduces a novel approach to routing that involves landmark-based routes and forwarder self-election, exploiting the knowledge of the underlying road network. We demonstrate TrafRoute's efficiency and robustness through simulation studies performed with accurate mobility and propagation models.

Simulation is essential to evaluate the performance of large scale vehicular networks. It is logistically challenging (and prohibitively expensive) to run tests with more than a few dozens experimental vehicles. Given the critical role of simulation in the evaluation of VANET protocols in large scale scenarios, it is important to guarantee realism of the models. This paper focuses on the accuracy of urban propagation models and their impact on vehicular protocol results. In a city-based vehicular network we compare the predominant two ray model and a recently proposed Corner model. We identify a number of factors that undermine the validity of the Two Ray model, for example, the presence of buildings causing propagation disruption and the heavy weight border effects that incorrectly compensate for the presence of hidden terminals in the networks. The paper analyzes a small scale urban vehicular scenario which unveils the issues to be considered in large scale vehicular simulations.