Wyniki wyszukiwania - Biblioteka Nauki

1

An adversarial explainable artificial intelligence (XAI) based approach for action forecasting

100%

Dutta V. , Zielińska T.

Journal of Automation Mobile Robotics and Intelligent Systems

|

2020

|

tom Vol. 14, No. 4

3--10

EN

Despite the growing popularity of machine learning technology, vision‐based action recognition/forecasting systems are seen as black‐boxes by the user. The effecti‐ veness of such systems depends on the machine learning algorithms, it is difficult (or impossible) to explain the de‐ cisions making processes to the users. In this context, an approach that offers the user understanding of these re‐ asoning models is significant. To do this, we present an Explainable Artificial Intelligence (XAI) based approach to action forecasting using structured database and object affordances definition. The structured database is sup‐ porting the prediction process. The method allows to vi‐ sualize the components of the structured database. Later, the components of the base are used for forecasting the nominally possible motion goals. The object affordance explicated by the probability functions supports the se‐ lection of possible motion goals. The presented methodo‐ logy allows satisfactory explanations of the reasoning be‐ hind the inference mechanism. Experimental evaluation was conducted using the WUT‐18 dataset, the efficiency of the presented solution was compared to the other ba‐ seline algorithms.

2

Improved competitive neural network for classification of human postures based on data from RGB-D sensors postures based on data from RGB-D sensors

80%

Dutta V. , Cydejko J. , Zielińska T.

|

2023

|

tom Vol. 17, No. 3

15--29

EN

The cognitive goal of this paper is to assess whether marker-less motion capture systems provide sufficient data to recognize human postures in the side view. The research goal is to develop a new posture classification method that allows for analysing human activities using data recorded by RGB‐D sensors. The method is insensitive to recorded activity duration and gives satisfactory results for the sagittal plane. An improved competitive Neural Network (cNN) was used. The method of pre- processing the data is first discussed. Then, a method for classifying human postures is presented. Finally, classification quality using various distance metrics is assessed. The data sets covering the selection of human activities have been created. Postures typical for these activities have been identified using the classifying neural network. The classification quality obtained using the proposed cNN network and two other popular neural networks were compared. The results confirmed the advantage of cNN network. The developed method makes it possible to recognize human postures by observing movement in the sagittal plane.

3

Klasyfikacja postur człowieka z wykorzystaniem sieci neuronowej

80%

Cydejko J. , Dutta V. , Zielińska T.

Prace Naukowe Politechniki Warszawskiej. Elektronika

|

2022

|

tom z. 197, t. 2

39-50

PL

Rozpoznawanie postur jest potrzebne do analizy czynności wykonywanych przez człowieka, syntezy ruchu robotów humanoidalnych oraz w badaniach nad robotami współpracującymi. Popularne w ostatnich latach tanie, bezprzewodowe czujniki wizyjne RGB-D umożliwiające łatwa˛ rejestracje˛ ruchu człowieka, ułatwiają˛ realizacje˛ tego zadania badawczego. Powszechnie stosowane są, tu konwencjonalne klasyfikatory, które na podstawie zarejestrowanych trajektorii ruchu stawów rozpoznają˛ postury człowieka. Nie sprawdzają, się, one jednak w pełni podczas obserwacji ruchu w płaszczyźnie strzałkowej, gdy pozycje niektórych stawów są, przesłonięte. Celem niniejszej pracy jest opracowanie nowej metody klasyfikacji postur. Opracowana metoda umożliwia rozpoznawanie aktywności z wykorzystaniem konkurencyjnej sieci neuronowej (CNN). Badania zrealizowano w dwu etapach. Pierwszy etap obejmował wstępne przetwarzanie danych, obejmujący filtrowanie i transformacje˛ danych. Drugim etapem było dobranie odpowiedniej sieci klasyfikującej. Zastosowano sieć konkurencyjną z uczeniem nienadzorowanym. Jakość klasyfikacji była testowana w zależności od różnych metryk odległości. Po wyborze satysfakcjonującej metryki dokonano jakościowej i ilościowej oceny uzyskanych wyników. Porównano wyniki klasyfikacji uzyskane przy użyciu rożnych typów sieci neuronowych. W zakończeniu pracy sformułowano szereg wniosków.

EN

Popular in recent years, cheap, wireless RGB-D vision sensors enable easy registration of human movement to facilitate the implementation of this research task. Here, the conventional classifiers are commonly used, which recognize human postures based on the recorded trajectories of joint movement. However, it does not perform well when observing movement in the sagittal plane because the positions of some joints are invisible here. The aim of the work is to develop a method for postures classification leading to the classification of human activities using data recorded by the RGB-D cameras. The method enables the recognition of activity with the use of a competitive neural network (CNN). The research was carried out in two stages. First, data preprocessing was performed, including raw data filtering and transformation to locate the origin of the reference frame in the point in the torso and scaling to obtain a consistent representation of the data. The classification quality was then tested using different variants of the proposed unsupervised variants of competitive neural networks. Various measures of distance have been used here. The results were presented graphically using scatter plots and stick diagrams.