Computer Science Journal (AGH University of Science and Technology, Krakow)
Not a member yet
476 research outputs found
Sort by
Immersive feedback in fencing training using mixed reality
Full text linkProviding athletes, during sports training, with real-time feedback based on automatic analysis of motion is both useful and challenging. In this work, a novel system based on mixed reality is proposed and verified. The system allows for immersive, real-time, visual feedback in fencing weapon practice. Novel methods are introduced for 3D blade tracking from a single RGB camera, creating weapon action models by recording actions performed by the coach and evaluating fencers\u27 performance against these models. Augmented reality glasses with see-through displays are employed and a method for coordinate mapping between virtual and real environments is proposed, which allows providing real-time visual cues and feedback by overlaying virtual trajectories on the real-world view. The system is verified experimentally in fencing bladework practice, with supervision of a fencing coach. Results indicate that the proposed system allows novice fencers to perform the exercises more correctly
Efficient multi-classifier wrapper feature-selection model: application for dimension reduction in credit scoring
Full text linkThe task of identifying most relevant features for a credit scoring application is a challenging task. Reducing the number of redundant and unwanted features is an inevitable task to improve the performance of the credit scoring model. The wrappers approach is usually used in credit scoring applications to identify the most relevant features. However, this approach suffers from the issue of subsets generation and the use of a single classifier as an evaluation function. The problem here is that each classifier may give different results which can be interpreted differently. Hence, we propose in this study an ensemble wrapper feature selection model which is based on a multi-classifiers combination. In a first stage, we address the problem of subsets generation by minimizing the search space through a customized heuristic. Then, a multi-classifier wrapper evaluation is applied using two classifier arrangement approaches in order to select a set of mutually approved set of relevant features. The proposed method is evaluated on four credit datasets and has shown a good performance compared to individual classifiers results
Group Membership Management Framework for Decentralized Collaborative Systems
Full text linkScientific and commercial endeavors could benefit from cross-organizational, decentralized collaboration, which becomes the key to innovation. This work addresses one of its challenges, namely efficient access control to assets for distributed data processing among autonomous data centers. We propose a group membership management framework dedicated for realizing access control in decentralized environments. Its novelty lies in a synergy of two concepts: a decentralized knowledge base and an incremental indexing scheme, both assuming a P2P architecture, where each peer retains autonomy and has full control over the choice of peers it cooperates with. The extent of exchanged information is reduced to the minimum required for user collaboration and assumes limited trust between peers. The indexing scheme is optimized for read-intensive scenarios by offering fast queries -- look-ups in precomputed indices. The index precomputation increases the complexity of update operations, but their performance is arguably sufficient for large organizations, as shown by conducted tests. We believe that our framework is a major contribution towards decentralized, cross-organizational collaboration
Set Representation for Rule Generation Algorithms
Full text linkThe task of mining the association rule has become one of the most widely used discovery pattern methods in Knowledge Discovery in Databases (KDD). One such task is to represent the itemset in the memory. The representation of the itemset largely depend on the type of data structure that is used for storing them. Computing the process of mining the association rule im- pacts the memory and time requirement of the itemset. With the increase in the dimensionality of data and datasets, mining such large volume of datasets will be difficult since all these itemsets cannot be placed in the main memory. As representation of an itemset greatly affects the efficiency of the rule mining association, a compact and compress representation of an itemset is needed. In this paper, a set representation is introduced which is more memory and cost efficient. Bitmap representation takes one byte for an element but the set representation uses one bit. The set representation is being incorporated in Apriori Algorithm. Set representation is also being tested for different rule generation algorithms. The complexities of these different rule generation algorithms using set representation are being compared in terms of memory and time execution
Plant disease detection using ensembled CNN framework
Full text linkAgriculture exhibits the prime driving force for growth of agro-based economies globally. In the field of agriculture, detecting and preventing crops from attacks of pests is the major concern in today\u27s world. Early detection of plant disease becomes necessary to prevent the degradation in the yield of crop production. In this paper, we propose an ensemble based Convolutional Neural Network (CNN) architecture that detects plant disease from the images of the leaves of the plant. The proposed architecture takes into account CNN architectures like VGG-19, ResNet-50, and InceptionV3 as its base models, and the prediction from these models is used as an input for our meta-model (Inception-ResNetV2). The approach helped us in building a generalized model for disease detection with an accuracy of 97.9 % under test conditions
The Impact of n-stage latent dirichlet allocation on analysis of headline classification
No full textData analysis becomes difficult with the increase of large amounts of data. More specifically, extracting meaningful insights from this vast amount of data and grouping them based on their shared features without human intervention requires advanced methodologies. There are topic modeling methods to overcome this problem in text analysis for downstream tasks, such as sentiment analysis, spam detection, and news classification. In this research, we benchmark several classifiers, namely Random Forest, AdaBoost, Naive Bayes, and Logistic Regression, using the classical LDA and n-stage LDA topic modeling methods for feature extraction in headlines classification. We run our experiments on 3 and 5 classes publicly available Turkish and English datasets. We demonstrate that n-stage LDA as a feature extractor obtains state-of-the-art performance for any downstream classifier. It should also be noted that Random Forest was the most successful algorithm for both datasets
Metadata-driven Data Migration from Object-relational Database to NoSQL Document-oriented Database
Full text linkThe object-relational databases (ORDB) are powerful for managing complex data, but they suffer from problems of scalability and managing large-scale data. Therefore, the importance of the migration of ORDB to NoSQL derives from the fact that the large volume of data can be handled in the best way with high scalability and availability. This paper reports our metadata-driven approach for the migration of the ORDB to document-oriented NoSQL database. Our data migration approach involves three major stages: a preprocessing stage, to extract the data and the schema\u27s components, a processing stage, to provide the data transformation, and a post-processing stage, to store the migrated data as BSON documents. The approach maintains the benefits of Oracle ORDB in NoSQL MongoDB by supporting integrity constraint checking. To validate our approach, we developed OR2DOD (Object Relational to Document-Oriented Databases) system, and the experimental results confirm the effectiveness of our proposal
Modeling and Analysis of Probabilistic Real-time Systems through Integrating Event-B and Probabilistic Model Checking
Full text linkEvent-B is a formal method used in the development of safety critical systems. However, these systems may introduce uncertainty, and need also to meet real-time requirements, which make their modeling and analysis a challenging task. Existing works on extending Event-B with probability and time did not address both probability and time in a single framework. Besides, they did focus the most on extending the language itself, not on integrating the extended Event-B with verification. In this paper, we aim to represent both probability and time in the Event-B language, and we will show how such a representation can be automatically translated into Probabilistic Timed Automata (PTA) described in the language of the probabilistic model checker PRISM. This translation would allow us to analyze probabilistic, as well as time-bounded probabilistic reachability properties of probabilistic real-time systems through the Probabilistic Timed CTL (PTCTL) logic
Dynamic Fuzzy Model to Detect Verbal Violence in Real Time
No full textThe crime rates in Mexico have been increasing in recent years, every day there are news on social media and in the news where assaults and verbal aggressions by criminals can be seen. Public transportation units suffer from violence that authorities have not been able to reduce, despite their efforts. That is why we have developed a fuzzy logic model that can adapt to almost any scenario thanks to the dynamism that we have implemented in each one of its stages. We have obtained promising results that we believe will be of great help to the authorities in the police headquarters to detect in real time the exact moment in which a verbal aggression typical of a violent assault is happening. This is a tool to help the authorities, not a substitution; making use of the latest technologies available to us
Human Gesture Recognition using Hidden Markov Models and Sensor Fusion
Full text linkConsidering the continued drive of human needs besides the constant improvement of technology, it is convenient to develop techniques that enhance the communication between computers and humans in the most intuitive ways as possible. The possibility to automatically recognize human gestures using artificial vision among other kind sensorsallows to explore a whole range of interaction applications to control and interact with environments. Nowadays, most of approaches for gesture recognition using sensors agree in the use of vision, myography and movement devices applied to robotic, medical and industrial applications. In the context of this work, we study the principles of using both vision andbody contact sensing applied to automatic classification of a human gesture set. For this, two different approaches are evaluated: Feed-forward Neural Networks and Hidden Markov Models. These models are studied and implemented for the recognition up to eight different human hand gestures commonly applied in collaborative robotics tasks. In our tests,we conclude the effectiveness of combining the information of two different sort of devices for human gesture recognition reaching accuracy rates up to 95.05% for a whole proposed Hand-gesture set