Action recognition in videos is currently a topic of interest in the area of computer vision, due to potential applications such as: multimedia indexing, surveillance in public spaces, among others. Attention mechanisms have become a very important concept within deep learning approach, their operation tries to imitate the visual capacity of people that allows them to focus their attention on relevant parts of a scene to extract important information. In this paper we propose a soft attention mechanism adapted to a base CNN–LSTM architecture. First, a VGG16 convolutional neural network extracts the features from the input video. Then an LSTM classifies the video into a particular class. To carry out the training and testing phases, we used the HMDB-51 and UCF-101 datasets. We evaluate the performance of our system using accuracy as an evaluation metric, obtaining 40,7 % (base approach), 51,2 % (with attention) for HMDB-51 and 75,8 % (base approach), 87,2 % (with attention) for UCF-101.

Action recognition in videos is currently a topic of interest in the area of computer vision, due to potential applications such as: multimedia indexing, surveillance in public spaces, among others. In this work we propose an attention mechanism adapted to a CNN–LSTM base architecture. To carry out the training and testing phases, we used the HMDB-51 and UCF-101 datasets. We evaluate the performance of our system using accuracy as the evaluation metric, obtaining 57.3% and 90.4% for HMDB-51 and UCF-101 respectively.

El reconocimiento de acciones en videos es actualmente un tema de interés en el área de la visión por computador, debido a potenciales aplicaciones como: indexación multimedia, vigilancia en espacios públicos, entre otras. Los mecanismos de atención se han convertido en un concepto muy importante dentro del enfoque de aprendizaje profundo, su operación intenta imitar la capacidad visual de las personas que les permite enfocar su atención en partes relevantes de una escena para extraer información importante. En este artículo proponemos un mecanismo de atención suave adaptado para degradar la arquitectura CNN–LSTM. Primero, una red neuronal convolucional VGG16 extrae las características del video de entrada. Para llevar a cabo las fases de entrenamiento y prueba, usamos los conjuntos de datos HMDB-51 y UCF-101. Evaluamos el desempeño de nuestro sistema usando la precisión como métrica de evaluación, obteniendo 40,7 % (enfoque base), 51,2 % (con atención) para HMDB-51 y 75,8 % (enfoque base), 87,2 % (con atención) para UCF-101.

espanolEl reconocimiento de acciones en videos es actualmente un tema de interes en el area de vision por computadora, debido a potenciales aplicaciones como: indexacion multimedia, vigilancia en espacios publicos, entre otras. En este articulo proponemos: (1) Implementar una arquitectura CNN–LSTM para esta tarea. Primero, una red neuronal convolucional VGG16 previamente entrenada extrae las caracteristicas del video de entrada. Luego, una capa LSTM determina la clase particular del video. (2) Estudiar como la cantidad de unidades LSTM afecta el rendimiento del sistema. Para llevar a cabo las fases de entrenamiento y prueba, utilizamos los conjuntos de datos KTH, UCF-11 y HMDB-51. (3) Evaluar el rendimiento de nuestro sistema utilizando la precision como metrica de evaluacion, dado el balance existente entre las clases de los conjuntos de datos. Obtenemos un 93%, 91% y 47% de precision respectivamente para cada conjunto de datos, mejorando los resultados del estado del arte para los primeros dos. Ademas de los resultados obtenidos, la principal contribucion de este trabajo yace en la evaluacion de diferentes arquitecturas CNN-LSTM para la tarea de reconocimiento de acciones. EnglishAction recognition in videos is currently a topic of interest in the area of computer vision, due to potential applications such as: multimedia indexing, surveillance in public spaces, among others. In this paper we propose (1) The implementation of a CNN–LSTM architecture. First, a pre-trained VGG16 convolutional neural network extracts the features of the input video. Then, an LSTM classifies the video sequence in a particular class. (2) A study of how the number of LSTM units affects the performance of the system. To carry out the training and test phases, we used the KTH, UCF-11 and HMDB-51 datasets. (3) An evaluation of the performance of our system using accuracy as evaluation metric, given the existing balance of the classes in the datasets. We obtain 93%, 91% and 47% accuracy respectively for each dataset, improving state of the art results for the former two. Besides the results attained, the main contribution of this work lays on the evaluation of different CNN-LSTM architectures for the action recognition task.

Action recognition in videos is currently a topic of interest in the area of computer vision, due to potential applications such as: multimedia indexing, surveillance in public spaces, among others. Attention mechanisms have become a very important concept within deep learning approach, their operation tries to imitate the visual capacity of people that allows them to focus their attention on relevant parts of a scene to extract important information. In this paper we propose a soft attention mechanism adapted to a base CNN–LSTM architecture. First, a VGG16 convolutional neural network extracts the features from the input video. Then an LSTM classifies the video into a particular class. To carry out the training and testing phases, we used the HMDB-51 and UCF-101 datasets. We evaluate the performance of our system using accuracy as an evaluation metric, obtaining 40.7% (base approach), 51.2% (with attention) for HMDB-51 and 75.8% (base approach), 87.2% (with attention) for UCF-101.

Action recognition in videos is currently a topic of interest in the area of computer vision, due to potential applications such as: multimedia indexing, surveillance in public spaces, among others. In this paper we propose (1) Implement a CNN–LSTM architecture. First, a pre-trained VGG16 convolutional neural network extracts the features of the input video. Then, an LSTM classifies the video in a particular class. (2) Study how the number of LSTM units affects the performance of the system. To carry out the training and test phases, we used the KTH, UCF-11 and HMDB-51 datasets. (3) Evaluate the performance of our system using accuracy as evaluation metric. We obtain 93%, 91% and 47% accuracy respectively for each dataset.

The automatic generation of video description is currently a topic of interest in computer vision due to applications such as web indexation, video description for people with visual disabilities, among others. In this work we present a Neural Network architecture Encoder-Decoder. First, a Convolutional Neural Network 3D extracts the features of the input video. Then, an Long Short-Term Memory decodes the vector to automatically generate the description of the video. To perform the training and testing we use the Microsoft Video Description Corpus data set (MSVD). Evaluate the performance of our system using the challenge of COCO Image Captioning Challenge. We obtain as results 0.3984, 0.2941 and 0.5052 for the BLEU, METEOR and CIDEr metrics respectively. Competitive results compared with certificates in the bibliography.

This paper describes a real-time feature-based stereo SLAM system that is robust and accurate in a wide variety of conditions - indoors, outdoors, with dynamic objects, changing light conditions, fast robot motions and large-scale loops. Our system follows a parallel-tracking-and-mapping strategy: a tracking thread estimates the camera pose at frame rate; and a mapping thread updates a keyframe-based map at a lower frequency. The stereo constraints of our system allow a robust initialization - avoiding the well-known bootstrapping problem in monocular systems-and the recovery of the real scale. Both aspects are essential for its practical use in real robotic systems that interact with the physical world.In this paper we provide the implementation details, an exhaustive evaluation of the system in public datasets and a comparison of most state-of-the-art feature detectors and descriptors on the presented system. For the benefit of the community, its code for ROS (Robot Operating System) has been released. (C) 2017 Elsevier B.V. All rights reserved.

Pedestrian detection is currently a topic of interest in computer vision due to its applications such as driver assistance systems and surveillance in public spaces, among others. The good results obtained using deep convolutional networks in vision tasks make them an attractive tool to improve the capacities of pedestrian detection systems. In this work we propose a deep convolutional network architecture to classify as pedestrian or non-pedestrian the candidate regions previously generated using a simple pyramidal sliding window approach. A distintive characteristic of the CNN in this system is that it separates pedestrian from non-pedestrian images without the aid of a pre-classification stage, and without the need of special tuning steps or initials conditions, making it more straightforward than other CNN-based solutions.The data used for training and testing come from the Caltech-USA Pedestrian dataset [8, 9]. We have evaluated the classification results on the proposed architecture and have obtained an average of ~ 98% success for the validation set. We have also carried out another evaluation of our system using the Benchmark proposed by Dollάr et.al. [9], and obtained results that are competitive with those mentioned in the bibliography.

Gender recognition is a topic of interest in computer vision due to its applications such as surveillance in public places, directed advertising, among others. The good results obtained using deep convolutional neural networks in vision tasks make them an attractive tool to improve the capacities of gender recognition systems. In this work we propose a deep convolutional network architecture to classify as male or female person the candidate regions previously detected using Haar features embedded in an AdaBoost. The data set used for training and testing come from the Labeled Faces in the Wildand Gallagher's dataset. We have evaluated the classification results on the proposed architecture and have obtained an average of ~95.42% and ~91.48% accuracy for the training set and for the test set, respectively, that are competitive with those mentioned in the bibliography. We have also carried out a real-time evaluation of the system using a web camera.

This paper describes a visual SLAM system based on stereo cameras and focused on real-time localization for mobile robots. To achieve this, it heavily exploits the parallel nature of the SLAM problem, separating the time-constrained pose estimation from less pressing matters such as map building and refinement tasks. On the other hand, the stereo setting allows to reconstruct a metric 3D map for each frame of stereo images, improving the accuracy of the mapping process with respect to monocular SLAM and avoiding the well-known bootstrapping problem. Also, the real scale of the environment is an essential feature for robots which have to interact with their surrounding workspace. A series of experiments, on-line on a robot as well as off-line with public datasets, are performed to validate the accuracy and real-time performance of the developed method.

Gender recognition is a relevant problem due to the number and importance of its possible application areas. The challenge is to achieve high recognition rates in the shortest possible time. Most studies are based on Local Binary Patterns (LBP) and its variants to estimate gender. In this paper, we propose the use of Binary Robust Independent Elementary Features (BRIEF), Oriented FAST and Rotated BRIEF (ORB) and Binary Robust Invariant Scalable Keypoints (BRISK) in gender recognition due to their good performance and speed. The aim is to show that ORB and BRISK are faster than LBP but allow to achieve similar recognition rates, which makes them suitable for real-time systems. For the best of our knowledge, it has not been studied in literature.

This work addresses the problem of motion segmentation in video sequences using dynamic textures. Motion can be globally modeled as a statistical visual process know as dynamic texture. Specifically, we use the mixtures of dynamic textures model which can simultaneously handle different visual processes. Nowadays, GPU are becoming increasingly popular in computer vision applications because of their cost-benefit ratio. However, GPU programming is not a trivial task and not all algorithms can be easily switched to GPU. In this paper, we made two implementations of a known motion segmentation algorithm based on mixtures of dynamic textures. One using CPU and the other ported to GPU. The performance analyses show the scenarios for which it is worthwhile to do the full GPU implementation of the motion segmentation process.

In this work an obstacle avoidance method is proposed for mobile robot autonomous navigation using a stereo camera as unique sensor. For each point of the world the disparity is computed from left and right images. With this approach, it is possible to determine the distance of nearby objects. Moreover, an elevation map is also computed for each pair of images. This map is used to remove the floor information in order to consider only the distances to the objects. Finally, using the depth and elevation maps an heuristic is defined to guide the robot avoiding obstacles. The method was successfully tested on indoor and outdoor environments.

We present an interactive easy-to-use software package, based on an evolutionary algorithm, to perform adaptive anisotropic diffusion speckle filtering for synthetic aperture radar (SAR) images. As a main difference from other methodologies, there is an integration of a SAR-image human expert who provides a subjective validation to complete the diffusion filter design. We have applied an interactive methodology to a set of SAR images, and we compared the results with those obtained by other speckle reduction filters. The results, through the evaluation of objective global and local quality criteria, show the potential of the proposal.

We present an approach for polarimetric Synthetic Aperture Radar (SAR) image region boundary detection based on the use of B-Spline active contours and a new model for polarimetric SAR data: the $${\mathcal{G}_P^H}$$ distribution. In order to detect the boundary of a region, initial B-Spline curves are specified, either automatically or manually, and the proposed algorithm uses a deformable contours technique to find the boundary. In doing this, the parameters of the polarimetric $${\mathcal{G}_P^H}$$ model for the data are estimated, in order to find the transition points between the region being segmented and the surrounding area. This is a local algorithm since it works only on the region to be segmented. Results of its performance are presented.

The human iris has proved to be one of the most reliable biometric features for the identification of individuals. Realtime iris recognition requires high quality images that provide enough details about the iris texture and algorithms to analyze and process the images at the highest possible speed. In this work, an extension to the classical circular model for the pupil and iris using flexible contours is provided. Then, a method for assessing the quality of the iris images in real-time based on the segmentation results is introduced. Experimental results are presented, and we conclude that the new methods improve the recognition rate, achieving a 100% correct recognition rate on the CASIA iris database, while being suitable for a real-time iris recognition camera system.

We compare the accuracy of five approaches for contour detection in speckled imagery. Some of these methods take advantage of the statistical properties of speckled data, and all of them employ active contours using B-spline curves. Images obtained with coherent illumination are affected by a noise called speckle, which is inherent to the imaging process. These data have been statistically modeled by a multiplicative model using the G0 distribution, under which regions with different degrees of roughness can be characterized by the value of a parameter. We use this information to find boundaries between regions with different textures. We propose and compare five strategies for boundary detection: three based on the data (maximum discontinuity on raw data, fractal dimension and maximum likelihood) and two based on estimates of the roughness parameter (maximum discontinuity and anisotropic smoothed roughness estimates). In order to compare these strategies, a Monte Carlo experience was performed to assess the accuracy of fitting a curve to a region. The probability of finding the correct edge with less than a specified error is estimated and used to compare the techniques. The two best procedures are then compared in terms of their computational cost and, finally, we show that the maximum likelihood approach on the raw data using the G0 law is the best technique.

Tracking objects in image sequences is a difficult problem because of the frequent occlusions encountered among the objects and people. In this paper, a method for tracking objects through occlusions based on a level set technique, is presented. Two cases are considered: 1-the object is obstructed by another one of a different color.2-the object is obstructed by another one of the same color.In the first case, the classical algorithms based on active contours fail to track the object's contour because the curve disappears from the scene. In this proposal, occlusion is detected and the solution is based on keeping the curve in the same place until the object of interest appears in the scene again.In the second case, the classic algorithms based on active contours do not distinguish between two image objects of the same color, so due to the possible changes in the topology, the curve fits the edge of both objects and it is divided. We solve this problem by means of localizing the centroid of the curve.The experimental results show that these methods significantly improve the tracking with occlusions.

Presently, the number of landmines planted around the world totalizes more than 110 million and, far from slowing down, the landmine production planting rate is, at least, one order of magnitude higher than the rate at which they are removed.In this work a technique to detect buried landmines using boundary detection in IR images, is presented. The buried objects have different temperature than the surrounding soil. We find the object contours by means of an algorithm of B-Spline deformable curves.Under a statistical model, regions with different temperatures can be characterized by the values of the statistical C, parameters of these distributions. Therefore, this information can be used to find boundaries among different regions in the image.The B-Spline approach has been widely used in curve representation for boundary detection, shape approximation, object tracking and contour detection. Contours formulated by means of B-Splines allow local control, require few parameters and arc intrinsically smooth. The algorithm consists in estimating the parameters along lines strategically disposed on the image. The true boundary is found when the values of these parameters vary abruptly on both sides. A likelihood function is maximized to determine the position of such boundaries.We present the experimental results, which show the behavior of the detection method, according to the buried object depth and the elapsed time from the cooling initial time. The obtained results exhibit that it is possible to recognize the shape of the objects, buried at different depths, with a low computational effort.