Abstract:
In the current technological context where signals can assist the functionality of the engines in operation and the correct functionality can be monitored. Therefore, patterns of utilization can be identified for predictive and preventive maintenance of such engines, thus predicting the Remaining Useful Life (RUL). For this reason, developing strategies to extract knowledge from recorded signals for preventing flaws is necessary and it opens an entire research direction. This paper presents the development of a generic strategy for exploring, analyzing and predicting the value of RUL and identifying techniques for specific data modeling. We defined and experimented a deep learning model, with a LSTM (Long Short-Term Memory) architecture. The identified strategies are tested and validated on a synthetic C-MAPSS data set which contains information from aircraft engines monitored and collected during several operating cycles. We defined 7 hypotheses, tested them and confirmed or unconfirmed each of them. We defined and presented: 6 architectural models, 3 sampling strategies on the original data set, presenting 18 representative experiments.

Abstract:
Exploring new areas for wind energy production brings new challenges for improvement of wind farms, making them more reliable and suitable for increasing of the power grids. In this regard, it is important to study and propose reliable solutions for contactless intelligent condition monitoring of the wind turbine blades. The article is intended to contribute to the study of various aspects of this actual multidisciplinary topic problem. In this context, the paper follows a systemic approach based on relevant general-purpose pragmatic quality criteria. It reports on the preliminary results obtained when analyzing the problem of building an embedded intelligent monitoring of the state of wind turbine blades using contactless strain sensors. For this reason, the numerical modeling of the blade deformations was performed in order to get the pattern of the maximum deformations of the blade. At the same time, from the pragmatic quality point of view, the required dataset, parameters of interest and intended data protocols were defined. Finally, a detailed structure of the edge computing module, as well as a preliminary framework for an embedded intelligent monitoring and decisionmaking system for predictive maintenance are presented.

Abstract:
Water supply infrastructure management has many challenging aspects that include smart infrastructure monitoring, data collection and storage, data analysis and decision support. In this paper we propose a prototype system that comprises solutions for monitoring water consumption, data collection and storage, as well as for data analysis, with the aim of providing support in the form of anomaly notifications and water consumption forecasts for residential consumers information applications.

Abstract:
The Digital Twins (DTs) offer promising solutions for smart grid challenges related to the optimal operation, management, and control of energy assets, for safe and reliable distribution of energy. These challenges are more pressing nowadays than ever due to the large-scale adoption of distributed renewable resources at the edge of the grid. DTs are leveraging on technologies such as Internet of Things (IoT), big data analytics, machine learning, and cloud computing, to analyze data from different energy sensors, view and verify the status of physical energy assets and extract useful information to predict and optimize the asset's performance. In this paper, we will provide an overview of the DTs application domains in the smart grid while analyzing existing the state-of-the-art literature. We have focused on the following application domains: energy asset modeling, fault and security diagnosis, operational optimization, and business models. Most of the relevant literature approaches found are published in the last 3 years showing that the domain of DTs application in smart grid is hot and gradually developing. Anyway, there is no unified view on the DTs implementation and integration with energy management processes, thus, much work still needs to be done to understand and automatize the smart grid management.

Abstract:
Medical image analysis is one of the most important applications in the fields of computer vision and medicine. Developing specific models that extract information from biomedical images could help doctors provide faster and more accurate diagnosis. One highly significant problem in this domain is called cell nuclei segmentation and it consists in identifying if a pixel from a medical image is a part of cell nucleus or not. Building such a model would help specialists detect several biomarkers of tumors. Nowadays, it is a well-known fact that the best tools for computer vision tasks are using deep neural networks, but they have one vulnerability - the data. Accurate and robust models rely on numerous labeled samples, but obtaining such a large, various annotated data set for cell nuclei segmentation is very difficult. One solution to tackle this problem is to use semisupervised learning which extends the data set with unlabeled samples during training. The aim of this paper is to build a model that performs cell nuclei segmentation in a semi-supervised manner by using Cross Pseudo-Supervision. We have run our experiments on 2018 Data Science Bowl dataset and we have achieved an IoU of 0.9077 using a fully-supervised setting, an IoU of 0.8493 using the same architecture, but in a semi-supervised setting, and an IoU of 0.7734 with a U-Net trained only on the labeled samples fed to the semi-supervised model. These values indicate the potential of machine learning, especially in those cases when only a small amount of labeled data are available.

Abstract:
Diabetic Retinopathy is a leading cause of irreversible blindness worldwide. However, early detection and timely treatment can significantly reduce morbidity and further vision loss. This work presents a lightweight automatic image processing / decision support system to detect early-stage diabetic retinopathy that would be suitable for implementation on a mobile device. In this study, we present a novel automatic and computationally uncomplicated extraction method to derive microaneurysms (a significant feature of early diabetic retinopathy). The proposed method is a sequence of morphological image processing operations including edge detection, boundary analysis, black top hat, image segmentation and logical operations. The work utilised images from the DIARETDB0, DIARETDB1, KAGGLE and MESSIDOR databases with classification further verified by a qualified ophthalmic diabetic specialist. The microaneurysm detection method returned an accuracy of 87%.

Abstract:
Eye diseases can nowadays be investigated with non-invasive OCT (Optical Coherence Tomography) tests which result in cross-sectional pictures of the retina. We use contrastive learning to learn a low-dimensional representation for OCT images which allows the reduction of OCT volumes to 2D images. The quality of the representation is analyzed from several perspectives. First, we analyze the ability to classify OCT B-scans with several eye lesions, i.e. drusen, choroidal neovascularization, and diabetic macular edema, respectively normal. Then, we shift to a particular disease, Age-related Macular Degeneration and give an answer to the question Is this representation capable to capture in a volume the aspects relevant to this disease and its effect on visual acuity despite the fact that it was not learnt from AMD dedicated data only?

Abstract:
Traditional wheat head detection and segmentation methods based on machine learning algorithms suffer from issues such as low efficiency and poor accuracy, resulting in the algorithms’ inability to generalize. The recent advances in deep learning, specifically in object detection methods, as well as computer development, have enabled the development of robust wheat head detection and segmentation methods. However, while international datasets of box labels are available for head detection, mask labels for segmentation are missing, and collecting them on a large scale is prohibitively expensive, time-consuming, and difficult. In this paper, we propose an unsupervised approach for segmenting wheat heads based only on box labels. Multiple state-of-the-art object detection methods have been trained on reference datasets and our collected data in order to find the best model to extract head bounding boxes. The obtained boxes were used as input of an unsupervised segmentation model named DeepMAC, which predicts the head mask in each box. Then, those masks are exploited to train several state-of-the-art supervised segmentation models. These models showed promising results on the collected dataset, covering all the wheat development stages. The average F1 score of head bounding box detection is 0.93 and the average F1 score of segmentation is 0.86.

Abstract:
In this paper we formulate an information retrieval problem as a sequence tagging task and develop and analyze critically various solutions. The specific context is that of extracting settings (such as cooking times, temperatures and cooking function) from online cooking recipes in multiple languages. We lack annotated data, therefore we start with developing a regex-based automatic tagger for three languages - English, German and French, which has proven to produce good enough annotations for learning. The final goal is to obtain a single extensible multilingual solution, and explore proven cross-lingual capabilities of pretrained multilingual language models to easily accommodate new languages. In the experimental evaluations performed we compare the performance of several monolingual and multilingual solutions, evaluate the impact of specific linguistic features on the performance of the models and assess zeroshot capabilities of the best multilingual solutions. The results indicate that a Conditional Random Field (CRF) layer added on top of Bidirectional Long-Short Term Memory (BiLSTM) models produces the best results in both mono- and multi-lingual settings, reaching F1-scores of over 96% for all classes, in both learning setups.

Abstract:
This paper addresses the problem of determining certain cooking parameters for monolingual recipes extracted from various websites. The problem is framed as multiple classification tasks, which we address with a number of (increasingly complex) solutions, considering various text representation alternatives (both sparse and dense) and several classifiers, investigating the effect of certain domain engineered features, exploring a joint learning strategy to exploit information sharing between the different classification problems and addressing specific challenges for class imbalance and varying text lengths. We found that textual information alone fed into domain adapted pretrained language models is enough to obtain the best classification accuracy. Moreover, a joint training approach, by uniformly adding the losses, significantly improves the accuracy on all classification tasks; performing a class-weighted loss aggregation does not further improve the behavior of the joint training approach.

Abstract:
News outlets employ catchy headlines that have the main purpose of luring the reader to click on the article and thus to visit the media site. These headlines are known as Clickbait. Although these baits may trick the readers into clicking, they often fail to deliver when it comes to the content of the article, leaving the reader disappointed. Furthermore, clickbaits can help the spread of misinformation, in the form of fake news, and have the power to influence people’s opinion. Our task is to automatically detect clickbaits from genuine article headlines. The developed system is based on deepforest, that is a non differentiable deep learning algorithm. As theoretical contribution, we extending the default deep-forest implementation with a boosting mechanism. The Chakraborty dataset was used that comprises 32,000 headlines of news articles. Three approaches were used for feature engineering, namely word embeddings, headline extracted features and a combination of both.

Abstract:
Keeping a journal can prove highly beneficial when it comes to confidence and overcoming psychological issues. There are plenty of applications that aim to help treat depression with the help of machine learning, using chat interfaces. However, little attention has been given to digital journals which take advantage of the latest AI technologies to help users better understand their emotions and track their evolution, for therapeutic purposes. A good understanding of human emotions is essential in the context of identifying, treating, and keeping track of psychological issues. There is extensive research in terms of emotion detection through video, audio, or pictures, and less when it comes to text input, despite the amount of data available publicly. We propose an application that, based on a model specialized in emotion detection, provides the user with meaningful insights which can be used to make informed decisions and take action.

Abstract:
witter Sentiment Analysis shows several challenges due to the platform’s features (e.g., short messages, colloquial style, etc.). People want to express their ideas related to personality, events, or breaking news. Social media is one of the fastest ways to express opinions, and research directions are developed to analyze the polarity of written messages. Very important for this domain is how you process the data and define features that can mine valuable information from textual inputs. Exploring various features and combining them can increase the quality of the entire methodology. Hence, a new system is designed to build a lexicon-based feature for detecting the polarity of a tweet. Therefore, messages posted by a user on Twitter are enhanced with a sentiment indicator provided by a lexicon. Then, the new model will be used by a classification algorithm. The numerical experiments are developed on several different datasets in terms of size and topics. The results highlight that the defined feature outperforms other lexicon-based features from literature. Moreover, the experiments based on the sentiment indicator produce better performance values than the traditional approaches that use the original tweet without additional features.

Abstract:
In this paper, we show how Federated Learning (FL) can be applied to vehicular use-cases in which we seek to classify obstacles, irregularities and pavement types on roads. Our proposed framework utilizes FL and TabNet, a state-ofthe-art neural network for tabular data. We are the first to demonstrate how TabNet can be integrated with FL. Moreover, we achieve a maximum test accuracy of 93.6%. Finally, we reason why FL is a suitable concept for this data set.

Abstract:
This article explores the utilization of the processing power of GPUs using CUDA computation for real-time aggregation of multi-sensor data and detection of 3D objects using parallel clustering algorithms. The purpose is to implement an algorithm that fuses raw lidar point cloud data and 2D camera image object detections to produce 3D object clusters in a lidar point cloud. Most of the computation has been implemented using CUDA parallelism to investigate the capability of GPU devices in this task, which is a common challenge in automated driving. The results indicate that processing times can be optimized within the algorithm, which is crucial when considering the large amounts of data provided by lidar and camera-based systems. The algorithm can perform inference on the Jetson Xavier AGX at rates of ~20 to ~220 ms depending on the number of objects and their corresponding point amounts in the KITTI dataset.

Abstract:
Efficiently encoding high definition maps is of great importance for autonomous navigation, thus they are widely used for tasks such as predicting the future behaviour of traffic participants and planning a safe trajectory. Previous methods tackled this problem either by rasterizing the road into a multichannel image, or by sampling the vectorial representation in fixed sized sub-segments (often called lanelets). The latter has become the go-to method due to its efficiency and expressiveness. However, its main limitation is the fact that the points creating a geometrical shape have to be sampled at a fixed spatial dimension, hence not taking full advantage of this representation's potential. In this work, we address this problem by making 2 additions to the classical architectures used for encoding such a heterogeneous structure. Rather than using a single network to encode a map element, we propose the decomposition of the map attributes such as road lines, traffic signs, road edges, etc. into structure and style features, an approach inspired by the recent progresses in the photo-realistic style transfer domain. The structural features will be encoded by a shared message passing network that treats the most essential positional data without the need of resampling at a fixed resolution, being able to adapt during inference the spatial dimension of the representation based on the initial length and complexity. The style attributes will be encoded separately and will allow for an easier addition of a new type of map element, without the retraining of the whole system. Evaluating the method on various edge prediction and node classification tasks proves that our method has better results than the previously mentioned approaches in both tasks, while having a 53% smaller memory footprint on average when representing a scenario.

Abstract:
An outdoor pedestrian-related risk estimation framework for traffic scenes using low cost visual sensors is proposed. The framework includes a Jetson Nano device with a monocular color camera for scene perception and processing. The risk estimation algorithm combines deep learning based approaches for road segmentation and object detection with conventional machine learning algorithms such as Random Forest, Support Vector Machines, Decision Trees, AdaBoost trained on visual and motion features computed for each frame. For evaluating the proposed algorithm a benchmark dataset for joint attention in autonomous driving and a custom dataset were annotated with the risk level each frame. The results show an accuracy of 90% for the benchmark data and 80% for the custom data series.

Abstract:
Recently, Intelligent Transportation Systems (ITS) have become one of the most important fields of research topics, while it provides advanced road scene monitoring. Actually, computer vision is one of the most widely used fields in ITS, while it offers various tasks, such as object detection, image classification, and segmentation. Besides, Convolutional Neural Networks (CNNs) have shown their effectiveness in deep learning tasks, particularly in computer vision. In fact, road scene semantic segmentation task is a critical and fateful issue that can be addressed using CNN which requires effective precision as well as fewer parameters. The majority of related work on road scene segmentation proposes models that focus on one aspect, the precision, or the parameters requirement, which makes it hard to use in real-time applications when the precision is not a priority. To solve this issue, we propose a new network based on an encoder-decoder architecture, which compromises the precision with fewer parameters. Our proposed model is trained from scratch using only 0.64M parameters. The experiments are evaluated on the popular CamVid dataset, and the results show that our proposed CNN achieves better performance with fewer parameter resources.

Abstract:
We compare here declarative approaches to model the LIGHTS OUT! game and its friends in First Order Logic (FOL). First, we solve the game using: (i) planning in FOL and (ii) a model finder for finite domains, for which we rely on Prover9 and Mace4. Second, we show how LIGHTS OUT! puzzles can be automatically generated by reasoning on finite models of FOL theories. We designed three solutions: (i) using a LIGHTS OUT! game solver in FOL, (ii) using linear algebra and a model generator, and (iii) improving the linear algebra-based method by decreasing the domain size. Third, we show how declarative knowledge can be reused to solve and generate different extensions of the LIGHTS OUT! game. We experimentally compare the proposed declarative methods and we discuss some extensions of theLIGHTS OUT! game.

Abstract:
Today, the huge amount of credit card transactions make detecting fraudulent ones increasingly difficult. This paper explores the idea of building a Neural Networks-based solution for detecting the fraudulent (i.e., unauthorized) transactions. The aim is to provide robust outputs and a reasonable trade-off be- tween Precision and Recall, while complying with the constraints imposed by working in a real time payment environment. The solution relies on a applying Genetic Programming to enhance the network performance. This is combined with the idea of a so-called ”Four-eyes Check” system, which tries to address the customer specific needs in terms of precision and recall. The approach was validated on a publicly available data set which contains real life credit card transactions on a span of two days . With no feature engineering except scaling, the system’s performance hit 80% on both precision and recall. Using the Four-eyes check lead to an increase of about 3%, depending on the specific functioning mode chosen.

Abstract:
In today's fast-paced, customer-focused marketing world, customer service management is key to revenue growth and profitability. Customer behavior knowledge can help marketing managers reevaluate existing customer tactics and plan to improve and increase effective strategies. Maintaining a productive relationship with business customers is crucial because business transactions require more decision-making and professional buying effort than consumer purchases. Most customer segmentation methods based on customer value fail to account for time and value changes. Today's business is based on new concepts because so many customers are unsure of what to buy. Businesses can't identify target customers. Machine learning algorithms detect hidden data patterns to make better decisions. Customer segmentation groups customers by gender, age, interests, and spending habits. Customers with unique needs are segmented by companies. Companies want to know their customers. Their goals must be clear and individualized. By analyzing data, businesses can better understand customer preferences and identify profitable segments. This helps them strategize their marketing while minimizing investment risk. Customer segmentation depends on many factors. Demographic, geographic, economic, and behavioural data help the company approach different sectors. Customer segmentation uses clustering to determine which consumer segment to target. This paper demonstrates machine learning-based customer segmentation. This is the unsupervised clustering problem, and we apply four of the most popular clustering algorithms: K-Means, Affinity Propagation, DBSCAN, and Hierarchical Clustering. We find the optimal number of clusters from a shopping mall dataset using the Elbow method and Silhouette score with the key factors like age, spending score, and annual income of customers. It helps shopping malls improve business by finding customer behavioural patterns. We compare the results using the Silhouette and DaviesBouldin scores. We find that Affinity Propagation and K-Means work best.

Abstract:
Massive Open Online Course (MOOC) provides online learning content to millions of students free of cost. But MOOC could not achieve their target due to a high rate of students not completing the courses. Students of MOOCs interact with teachers and other students through discussion forums of MOOC websites. Our idea is to collect comments from the MOOC website and predict user behaviour. We want to predict if any student needs help. We manually label each statement with 'yes' or 'no.' Using those labelled datasets, we train a machine to predict if an instructor intervention is needed to help students. Our model resulted in up to 72% accuracy using J48 algorithm. Our model works better with an increase in instances compared to other algorithms such as Naive Bayes.

Abstract:
Localization systems are crucial to every intelligent mobile system, especially to automated vehicles. Knowing the own position, together with sensor data and predefined information, is the basis for all driving decisions. Significant effort has been made up to this point to improve robustness and functionality of localization systems. Nonetheless, such safety critical components must not fail. Besides self-testing of a localization system another possibility is to externally observe it’s functionality. Possible failures might be error-prone pose measurements or timing issues. Although localization systems usually suffer from tiny inaccuracies, e.g. due to sensor noise and model assumptions, these do not affect the capabilities of the automated system in a critical way. Therefore, this paper presents a novel method for deriving the functionality of an arbitrary localization system by utilizing odometry data of the vehicle and comparing the pose and time differences to respective parameterized probability distributions learned in advance from error-free mapping drives. Since the performance of the localization system is positiondependent the method utilizes a grid map and calculates the best fitting probability distribution for each cell during the mapping phase. An evaluation of the proposed failure detection shows that it is capable of recognizing errors of the localization system.

Abstract:
We have previously proposed a linear approach for reducing the global drift of a video-based frame-to-frame trajectory estimation method by correcting it at selected points in time based on the alignment of past and current 3D LiDAR measurements (see [7]). In this paper we assess the tolerance to noise of a series of methods derived from the one previously proposed, this time using both linear and non-linear optimization methods to calculate the correction transform. We generate synthetic datasets with various noise pollution levels and assess the performance of each method under investigation in recovering artificially induced odometry estimation errors.

Abstract:
We present aSDF, a continual learning system based on attention for real-time signed distance field reconstruction. Given a stream of posed monocular RGB images from a moving camera, it trains a randomly initialized network to map input 3D coordinates to an approximated signed distance. The model is based on a transformer encoder enhanced with a hash table which acts as a cache, significantly improving the learning time, a crucial component in a continual learning system. In contrast to prior works, our model further improves the compact representation, while maintaining a decent accuracy in terms of the reconstructed signed distances. By employing self-supervised losses with a bigger penalization in the case of wrong prediction, the network is able to improve the accuracy of unobserved regions by filling out with a more plausible signed value. In evaluation to other methods on indoor synthetic and real datasets of indoor rooms, aSDF produces competitive reconstruction with previous state-of-the-art depth based methods and better approximation of gradients, while converging much faster due to the hashing component.

Abstract:
Deep learning techniques for machine learning usually lead to increasingly complex models while enhancing performance, which convert these systems into "black box" approaches and create confusion about how they function and, ultimately, how they make decisions. Layerwise Relevance Propagation, an Explainable AI (XAI) technique, excels at explaining any neural network's output in the context of its input by calculating relevance iteratively from the output class neuron to the original input neurons. This paper outlines recent advancements in the area and advocates for more interpretability in AI. In this paper, we study in-depth Layerwise Relevance Propagation (LRP) and about its four different LRP approaches, which we have experimented on different dataset. With the aim of explaining predictions of deep learning models in the field of image classification in this paper we have tried to basically understand how this different LRP approach helps in describing which pixels in an image are important for deciding on a classification decision by computing a relevance score on its given input image. Furthermore, in this paper, we did human eye evaluation by conducting a user questionnaire survey on XAI in order to compare the quality of heatmaps generated by different LRP methods.

Abstract:
Pavement surface condition rating is an essential part of road infrastructure maintenance and asset management, and it is performed manually by the data analyst. The manual rating requires cognitive skills built through training and experience, which is quantitatively challenging and timeconsuming. This paper first analyses the complexity of the current manual visual rating system. This paper then investigates the suitability and robustness of a state-of-the-art convolutional neural network (CNN) classifier to automate the pavement surface condition index (PSCI) system used to rate pavement surfaces in Ireland. The dataset contains 3735 images of flexible asphalt pavements from Irish urban and rural environments taken from a video camera mounted in front of a van. The PSCI ratings were applied by experts using a scale of 1-10 to indicate surface conditions. The classification models are evaluated for different input pre-processing variations, image size, learning techniques, and the number of classes. Using 10 PSCI classes, the best classifier achieved a precision of 57% and a recall of 58%. Adjacent combination of classes (e.g., ratings 1 and 2 combined into a single class) to form a 5-class problem produced a classifier with a precision of 70% and recall of 77%. Given the complexity of the problem, classification using CNN holds promise as a first step towards an automated ranking system.

Abstract:
In the latest years, several approaches for cognitive analysis were developed. One of those approaches addresses the perspective regarding the brain as a network. This paper tries to prove the utility of brain analysis using functional networks. For this experiment, we considered signals provided by DEAP dataset, on which we applied frequency filters and statistical methods for obtaining a graph-based model of the brain, for every subject and trial. A real challenge was the choice of the threshold that differentiates the relevance of the functional links between groups of neurons. Furthermore, in the attempt to prove the effectiveness of the considered method, the graph metrics were computed (average degree, betweenness centrality, global efficiency, cluster coefficients, etc.) and used for emotional state detection and classification. For the same classification problem, the classic features extracted from the brain waves (power spectrum, spectral entropy, Hjorth parameters, etc.) were considered. A comparison between the accuracy obtained from both classifiers is performed for the sole purpose of enforcing the utility of functional network driven methods in further analyses. Finally, by adding the graph metrics to the standard ones, better accuracy of the classification was obtained.

Abstract:
Correct depth of anesthesia (DOA) assessment is a serious and widespread medical problem and an active scientific research topic. Anesthesia is known to alter the dynamics of neural networks within the brain that ultimately results in the impairment of pain perception. Nevertheless, linking the composition and dosage of various anesthetics, to a target level of anesthesia is not trivial. In this paper we explore the use of microstates as a viable tool to discriminate between anesthesia levels in intracranial recordings from mouse visual cortex. We show such symbolic analysis is able to capture DOA specific information in local field potentials (LFP). Microstates are characterized by the appearance of set of prototypical maps of activations over recording sites (electrodes) that define stereotypical patterns of activation in the recorded neural networks. Although microstates have been defined and characterized in electroencephalogram (EEG) data, our study shows for the first time that microstates can be effective when considering LFPs as well. We performed statistical analysis of average duration, time coverage, and occurrence of the microstates in order to differentiate between different DOA levels. By increasing the number of microstates, the analysis is more insightful and it is easier to discriminate between DOA levels.

Abstract:
Because of the intricacy of neural data, spike sorting is a challenging problem in neuroscience. Overlapping clusters are one of the hardest issues to solve, along with the similarity of spike shapes, excessive noise, and unbalanced clusters. Here, we focus on analyzing the efficacy and integrating the Superlets Transform as a feature extraction method. Our results indicate that Superlets achieve not only super-resolution in time and frequency, but also provide the conditions for smooth clustering performance. The Superlet Transform was analyzed in conjunction with Principal Component Analysis, Singular Value Decomposition, and Isomap Embedding as dimensionality reduction methods. The Superlet-extracted features were classified using Neural Networks in order to assess their relevance.

Abstract:
Due to technological advances, massive amount of data is recorded in all areas. Not only the manufacturing processes are monitored and can benefit from data augmentation, but also the utilisation of products produce large amounts of information rich data. With the right set of skills and processing strategies, valuable process-related information can be extracted from such data. However, before applying any complex processing tasks such as user profiling, raw data needs to be preprocessed, and knowledge needs to be extracted. In this paper, we propose a set of steps for knowledge extraction from home appliances data, focusing on user and appliance specific data.

Abstract:
Nowadays, many companies are moving their activities to the cloud. They are using advantages such as collaboration, disaster recovery, and cost savings as the drive for this change. Nonetheless, many cloud providers have different offerings, some more appealing than others. This article describes decisive factors for the deployment of modeling environments using a multicloud deployment platform. This is conceived for users who are specialists in modeling, but not in cloud computing, but who intend to take advantage of cloud environments’ capabilities. The automate decisions are made by leveraging the decision trees and providing a decisional algorithm that helps one decide the deployment details. Furthermore, the paper evaluates the algorithm from the perspective of the most valuable outcome of it.

Abstract:
Smart building monitoring is prevalent today, having strong support from Cloud Computing technologies. Nonetheless, the effectiveness in doing the necessary processing in due time also depends on the devices situated on the edge, for acquiring data, and sometimes for making the first decision recommendations. The paper investigates a set of solutions with various implementations for the edge nodes and two ways of deploying the monitoring software – locally, inside the building, or in a public cloud. They were tested for real-world settings in a building from our university, to compare the response times and discuss if inherent cost differences are compensated by the performance achieved.

Abstract:
The FreeBSD hypervisor, bhyve, received a series of improvements over the years. Work was done on multiple disk support and improvements have been made to the state saving file format and the hypervisor security. One important functionality that is missing from the hypervisor is the checkpoint mechanism, which is needed by every robust hypervisor in order to be used at a larger scale. For bhyve, the article displays how existing functionalities can be joined to create an usable checkpoint mechanism that can serve as the base for a mature checkpoint implementation.

Abstract:
A technical university needs high performance computing resources to support the research activity carried out by interdisciplinary teams, through heterogeneous, various and flexible configurations. The CloudUT infrastructure has been designed and developed to meet the requirements and performances necessary for research projects in the scientific and engineering fields specific to the Technical University of Cluj-Napoca such as big data, artificial intelligence, spatial data, satellite data, internet of things, complex simulation, and computer assisted design. This paper analysis and highlight the challenges and issues on finding efficient solutions for the distributed architecture, parallel computing, massive data, flexible architectures, virtualization, cloud applications and services for covering a large range of technical domains of engineering such as mechanics, automation and computers, civil and building services, electric, architecture and urban planning, electronics, robotics, and mechatronics.

Abstract:
In this work we aim to construct a robust pipeline for the sports analysis community in order to successfully extract useful information from broadcast football matches. We propose a fast and efficient solution, based on computer vision and machine learning methods and algorithms. Our solution provides a framework suitable not only for detecting, tracking and identify the roles of the players and staff, but also for mapping each player from their position as seen in broadcast images, to their absolute position on the field. In order to achieve this, we designed each module of the pipeline by comparing multiple solutions and choosing the most suitable ones taking into consideration the trade-off between performance and inference time. We managed to provide a system that can be used by anybody in the community by feeding a sequence of frames taken from a broadcast football match.

Abstract:
The popularity of multiplayer games has risen in the recent years. As a result, digital items that are obtainable in such games have become more valuable and therefore the presence of malicious users that use software agents has become more prominent. The proposed framework aims to detect such bots in online games by identifying anomalies in the economy of the digital world and in their behavior. It aids the game managers with detailed reports on possible malicious user accounts and offers traceability of the digital items or currency generated.

Abstract:
Reverse engineering binary applications is a key process for black-box security auditing and malware analysis. Frida is a reverse engineering framework based on dynamic binary instrumentation that allows the user to create agents, which are injected in the analyzed process, and can communicate with the user’s program. Frida is written in C and Vala and offers high level bindings in Python and JavaScript. Dynamic languages allow fast development iteration, a key requirement when trying to discover the inner workings of an application or protocol. The main disadvantages of such languages include performance limitations and their error-prone nature due to lack of type checking. In this paper we address these limitations by building bindings in Rust, which aims to offer high performance and numerous correctness guarantees while still maintaining reasonable development iteration speed. We show examples of performance improvements and present a real use case to validate the usability of the library.

Abstract:
We investigate the potential of U-Net in performing knowledge distillation, both in the posture of Teacher and Student network. The experiments evaluate multiple configurations of the distillation procedure, as well as various Teacher-Student pairs, in order to exploit the capabilities of the architecture, as it is rarely used in this situation. The study was performed by applying the standard distillation process (standard KD), a pixel-wise approach of transferring Teacher predictions to the Student, as well as an Ensemble Teacher method, whose mean prediction is utilized in the training process of the student. These methods improved the results of the Student models to an extent worthy of being taken into consideration.

Abstract:
In this paper, we propose an approach for computing the 3D trajectory of UAVs between two locations when obstacles are present. The result is an obstruction free path, close to optimal. First, the precomputed 3D environment map is converted into a discrete voxel space. Next, the A* algorithm is applied on the discretized space. The A* method provides the optimal path with respect to the graph equivalent representation of the voxel space, however in the continuous 3D space the resulting path will cause unnecessary steering maneuvers for the UAV. The solution is to smooth the resulting path using an iterative linear approximation approach. A new representation of the 3D path is obtained, consisting of line segments and control points. Therefore, we manage to transform the A* path from the graph equivalent representation of the voxel space to the continuous representation of the 3D environment. The resulting control points can be used as intermediate destinations during autonomous UAV navigation. Experiments are performed on multiple scenarios, demonstrating that the proposed method shortens the standard A* path for UAV navigation in the 3D environment.

Abstract:
Humans use Unmanned Aerial Vehicles (UAVs) for a large variety of tasks like cinematography, agriculture, traffic monitoring, and disaster response. Before they are deployed in their missions, their functionalities can be tested in virtual scenarios, using simulators. These are programs that model the environment and physics of the real world, and offer a safe space for implementing and testing algorithms. In this paper, we propose the usage of AirSim and the creation of a natural environment in Unreal Engine 4 to solve the problem of lowaltitude autonomous drone navigation for forest monitoring. We analyze the performance of our navigation methodology in a virtual forest environment. Our results showcase the high level of realism that can be achieved by the simulator and the difficulties which arise when deploying a drone in complex conditions.

Abstract:
Deep learning-based solutions for the ill-posed problem of Monocular Depth Estimation (MDE) from 2D color images have shown potential in recent years, spurring a very active field of research. Most state-of-the-art proposals focus on solving the problem in the context of automotive advanced driver assistance and/or autonomous driving systems. While presenting their own complexities and challenges, the vast majority of road environments exhibit a number of commonalities amongst themselves. The aerial domain in which modern Unmanned Aerial Vehicles (UAVs) operate is significantly different and features a large variety of possible scenes based on the specific mission carried out. The increasing number of applications for UAVs could benefit from more advanced learning-based MDE solutions for recovering 3D geometric information from the scene. In this paper, we conduct a study of existing research on the topic of MDE specifically tailored for aerial views, as well as presenting the datasets and tools currently supporting such research, highlighting the challenges that remain. To the best of our knowledge, this is the first survey covering this field.

Abstract:
This study covers the avionics system design for an unmanned combat aerial vehicle (UCAV) that can successfully detect and lock down rival unmanned aerial vehicle targets as many times as possible in air-to-air combat. With the artificial intelligence (AI) supported development board and high resolution/wide-angle camera on it, easy detection of competitor systems is provided. The captured images and flight data are processed with the AI-algorithm working on the developer kit with 48 tensor cores on it. Moreover, flight telemetry data is transmitted to the ground station at a range of 40 km, end-to-end encrypted and with low latency. Through 2.4 GHz radio frequency control, manual control can also be performed in an encrypted way and at a sufficient distance. Designed as a hexacopter, the UCAV can also avoid locking up of counter-unmanned systems with its high maneuverability. As a result of the proper design of the avionics and electrical system architecture, it can perform different tasks such as snapshot and data transfer, as well as being able to fly fully autonomously.