The emerging branch of micro aerial vehicles (MAVs) has attracted a great interest for their indoor navigation capabilities, but they require a high quality video for tele-operated or autonomous tasks. A common problem of on-board video quality is the effect of undesired movements, so different approaches solve it with both mechanical stabilizers or video stabilizer software. Very few video stabilizer algorithms in the literature can be applied in real-time but they do not discriminate at all between intentional movements of the tele-operator and undesired ones. In this paper, a novel technique is introduced for real-time video stabilization with low computational cost, without generating false movements or decreasing the performance of the stabilized video sequence. Our proposal uses a combination of geometric transformations and outliers rejection to obtain a robust inter-frame motion estimation, and a Kalman filter based on an ANN learned model of the MAV that includes the control action for motion intention estimation.
In this article, we propose some methods for deriving symbolic interpretation of data in the form of rule based learning systems by using Support Vector Machines (SVM). First, Radial Basis Function Neural Networks (RBFNN) learning techniques are explored, as is usual in the literature, since the local nature of this paradigm makes it a suitable platform for performing rule extraction. By using support vectors from a learned SVM it is possible in our approach to use any standard Radial Basis Function (RBF) learning technique for the rule extraction, whilst avoiding the overlapping between classes problem. We will show that merging node centers and support vectors explanation rules can be obtained in the form of ellipsoids and hyper-rectangles. Next, in a dual form, following the framework developed for RBFNN, we construct an algorithm for SVM. Taking SVM as the main paradigm, geometry in the input space is defined from a combination of support vectors and prototype vectors obtained from any clustering algorithm. Finally, randomness associated with clustering algorithms or RBF learning is avoided by using only a learned SVM to define the geometry of the studied region. The results obtained from a certain number of experiments on benchmarks in different domains are also given, leading to a conclusion on the viability of our proposal.
Premi extraordinari ex-aequo en l'àmbit d'Electrònica i Telecomunicacions. Convocatoria 1999 - 2000
Nearest Neighbour (NN) classifiers are one of the most celebrated algorithms in machine learning. In recent years, interest in these methods has flourished again in several fields (including statistics, machine learning and pattern recognition) since, in spite of their simplicity, they reveal as powerful non-parametric classification systems in real-world problems. The present work is mainly devoted to the development of new learning algorithms for these classifiers and is focused on the following topics:
- Development of learning algorithms for crisp and soft k-NN classifiers with large margin - Extension and generalization of Kohonen's LVQ algorithms - Local stabilization techniques for ensembles of NN classifiers - Study of the finite-sample convergence of the on-line LVQ1 and k-means algorithms
Besides, a novel oriented principal component analysis (OPCA) addressed for feature extraction in classification is introduced. The method integrates the feature extraction into the classifier and performs global training to extract those features useful for the classifier. The application of this general technique in the context of NN classifiers derives in a problem of learning their weight metric.
Regularisation is a well-known technique for working with ill-posed and ill-conditioned problems that have been explored in a variety of different areas, including Bayesian inference, functional analysis, optimisation, numerical analysis and connectionist systems. In this paper we present the equivalence between the Bayesian approach to the regularisation theory and the Tikhonov regularisation into the function approximation theory framework, when radial basis functions networks are employed. This equivalence can be used to avoid expensive calculations when regularisation techniques are employed.