Humberto Loaiza

Matching segments in stereoscopic vision

We have shown that its possible to realize a stereoscopic sensor with poor cameras. We developed image processing that is robust and allows us to quickly obtain results for the matching algorithm. We computed an important number of features on each segment, and with these features, we built 16-component vector used in the classification step. After an exhaustive study, we decided to combine two methods, Bayesian and neural, to construct an efficient classifier. The tests for indoor images had better than 90% good matching. With segment couples, it is possible to compute the 3D coordinates of the objects. Therefore, the mobile robot is able to localize and move about in the environment.

Phase contrast using a differentiated absolute contrast method

A depth retrieval technique based on phase contrast calculations by pulsed phase thermography (PPT) has been previously reported. The phase contrast requires an appropriate selection of the sound area. This is rarely an easy task primarily because a non-defective zone is not always a priori known and even when it is, some variability is typically observed in the results due to changes in the sound phase. This article proposes implementing the differentiated absolute contrast (DAC) method to eliminate the need of defining a sound area. The proposed PPT-DAC approach, allows computation of the phase contrast by subtracting the ideal phase value of a pixel from its measured phase.

Procesamiento de imágenes infrarrojas para la detección de defectos en materiales

El Ensayo Termografico No Destructivo (ETND) es una tecnica de evaluacion de materiales, en la que la superficie de una muestra de material es estimulada tecnicamente para producir una diferencia de temperatura entre las areas no defectuosas y las eventualmente defectuosas. Los cambios de temperatura son registrados mediante una camara infrarroja; posteriormente, dada la distorsion generada por el ruido, se ejecutan etapas de procesamiento para detectar y/o caracterizar los defectos en el material. En este articulo se analizan y comparan experimentalmente varios de estos metodos de procesamiento y se profundiza en la tecnica CAD (Contraste Absoluto Diferencial) modificada por cuadrupolos termicos.

Defect quantification with reference-free thermal contrast and artificial neural networks

The Infrared Nondestructive Testing (IRNT) methods based on thermal contrast are strongly affected by non-uniform heating at the surface. Hence, the results obtained from these methods considerably depend on the chosen reference point. One of these methods is Artificial Neural Networks (ANN) that uses thermal contrast curves as input data for training and test in order to detect and estimate defect depth. The Differential Absolute Contrast (DAC) has been successfully used as an alternative thermal contrast to eliminate the need of a reference point by defining the thermal contrast with respect to an ideal sound area. The DAC technique has been proven effective to inspect materials at early times since it is based on the 1D solution of the Fourier equation. A modified DAC version using thermal quadrupoles explicitly includes the sample thickness in the solution, extending in this way the range of validity when the heat front approaches the sample rear face. We propose to use ANN to detect and quantify defects in composite materials using data extracted from the modified DAC with thermal quadrupoles in order to decrease the non-uniform heating and plate shape impact on the inspection.

Neural and statistical classifiers. Can such approaches be complementary

Neural networks are efficient in certain pattern recognition sub-problems, especially in feature extraction and classification. In many cases neural and statistical techniques are seen as alternatives. Our aim is to verify if these approaches can give complementary responses in order to consider the implementation of fusion methods. The comparison is applied to three examples belonging to mobile robot localization: (i) laser range finder modeling, (ii) feature extraction from ultrasonic range finder data and (iii) localization by a stereoscopic camera. In each case the solution of the problem is based partly on a classifier. The paper compares the performances of a multilayer perceptron (MLP) known as an efficient classifier and three statistical methods-quadratic discriminant analysis (QDA), linear discriminant analysis (LDA) and Bayesian. The performances of the classifier are estimated by classical criteria such as success and misclassification percentages and the study is completed by a sharp analysis where the method results are crossed two by two to evaluate the success percentage of a method applied to the misclassified set of another one. Experiments show the set of patterns misclassified by the different classifiers does not completely overlap.

Modelo bio-inspirado para el reconocimiento de gestos usando primitivas de movimiento en visión

This article addresses the issue of gesture recognition using movement primitives to obtain a bio-model that, in a close future, can be used in the robot programming through the imitation learning paradigm. Those movement primitives are extracted from consecutive images caught by a standard web cam. For robot programming by imitation, gesture recognition was identified as first phase, which requires three main aspects to be taken into consideration. These are the instantaneous movement representation, the temporal integration of related information, and the classification strategy. These three aspects are going to be developed in this article and in contrast to other works in this field; the movement extraction and its codification are inspired in the macaco´s brain motion processing. The obtained model was applied then to the recognition of four different hand gestures performed by different people. The success percentage using different standard classification strategies varied between 91.42% and 97.14%.

Identifying facial gestures to emulate a mouse: navigation application on Facebook.

A system is presented for emulate a mouse from the movement of the head and eyelids. The position of the head is used for controlling the horizontal and vertical displacement of the cursor, and the closing of the eyelids to activate the click of the right and left buttons. The system includes zoom, navigation shortcuts, vertical scrollbars and menus activation according to the cursor position; that enhance the functionality and facilitate handling applications. The proposed solution eliminates the restriction of direct contact with the mouse, and empowers people with motor disabilities in the upper extremities to interact with the computer. The mouse emulator can also be used by people without limitations to expand command instructions. The system was tested in navigation on social network Facebook, where an average speed of 382 pixels / s was obtained, with an average accuracy of 22 pixels for the X axis and 17 pixels for the Y axis. However, after user interaction with the interface, improvements of 23% and 37% were observed, in execution time and location accuracy, respectively. Click activation by temporary location over a menu option had a performance of 100%; while the right and left clicks by eyelid closing had a performance of 93% and 92%, respectively. Finally, the surveys showed high satisfaction about the proposed interface during user interaction with Facebook.

Identifying facial gestures to emulate a mouse: Control application in a web browser

Computer interfaces continue using mouse and keyboard, although the form of human interaction is based primarily on exchange of audiovisual messages generated by the voice and body gestures in general. Because of the importance of gestures in human communication and the possibility for people with limited mobility of the upper limbs, this research focuses on the development of an interface to emulate the mouse and keyboard through facial gestures and head movement. The system was developed to interact with a web browser by processing and analysis of images provided by a webcam. The interface allows typing for text searching, move the mouse, click by closing one eye, open and navigate between tabs, vertical scroll, zoom, back and leave. Tests show a moving speed of cursor 411 px/s and a write speed of 3.5 characters/s, correct activation of right and left clicks 94% and 92% respectively. Finally, the survey showed a high user satisfaction while browsing on Internet.

Locally Adapted Gain Control for Reliable Foreground Detection

One of the first steps in video analysis systems is the detection of objects moving in the scene, namely the foreground detection. Therefore, the accuracy and precision obtained in this phase have a strong impact on the performance of the whole system. Many camera manufacturers include internal systems, such as the automatic gain control (AGC), so as to improve the image quality; although some of these options enhance the human perception, they may also introduce sudden changes in the intensity of the overall image, which risk to be wrongly interpreted as moving objects by traditional foreground detection algorithms. In this paper we propose a method able to detect the changes introduced by the AGC, and properly manage them, so as to minimize their impact on the foreground detection algorithms. The experimentation has been carried out over a wide and publicly available dataset by adopted one well known background subtraction technique and the obtained results confirm the effectiveness of the proposed approach.

Classical and neural models for binocular stereoscopic reconstruction

We present a comparative study about stereoscopic reconstruction process focused in modelling for parallel axis stereoscopic cameras. We used two classical models and one based on Artificial Neural Networks for modelling the parallel axis system. Then we used the root mean square of distances between the point coordinates calculated from images and measured from the calibration pattern to evaluate the accuracy of each model. We compared the accuracy of two classical models and one based on Artificial Neural Networks. By comparing the confidence interval for every obtained model we observed that the classical model of Silven and Heikkila [1] shows average errors of 1.0 cm, however this error was reduced to 0.4 cm by an adjustment proposed in this paper. On the other hand the neural networks showed less robust to the training set. The current work can be extended to future developments in areas like photogrammetry, architecture and robotics.