Norberto A. Goussies

Detecting pedestrians on a Movement Feature Space

This work aims at detecting pedestrians in surveillance video sequences. A pre-processing step detects motion regions on the image using a scene background model based on level lines, which generates a Movement Feature Space, and a family of oriented histogram descriptors. A cascade of boosted classifiers generates pedestrian hypotheses using this feature space. Then, a linear Support Vector Machine validates the hypotheses that are likeliest to contain a person. The combination of the three detection phases reduces false positives, preserving the majority of pedestrians. The system tests conducted in our dataset, which contain low-resolution pedestrians, achieved a maximum performance of 25.5% miss rate with a rate of 10 - 1 false positives per image. This value is comparable to the best detection values for this kind of images. In addition, the processing time is between 2 and 6fps on 640×480 pixel captures. This is therefore a fast and reliable pedestrian detector. HighlightsA Movement Feature Space generates oriented histogram family descriptors.Detection system involves motion detection, hypothesis generation and validation.A cascade of classifiers combine generative and discriminants functions.Best result gives 25.5% of miss rate with 0.1 false positives per image.Running time is between 2 and 6fps in 640×480 frames size.

Transfer learning decision forests for gesture recognition

Decision forests are an increasingly popular tool in computer vision problems. Their advantages include high computational efficiency, state-of-the-art accuracy and multi-class support. In this paper, we present a novel method for transfer learning which uses decision forests, and we apply it to recognize gestures and characters. We introduce two mechanisms into the decision forest framework in order to transfer knowledge from the source tasks to a given target task. The first one is mixed information gain, which is a databased regularizer. The second one is label propagation, which infers the manifold structure of the feature space. We show that both of them are important to achieve higher accuracy. Our experiments demonstrate improvements over traditional decision forests in the ChaLearn Gesture Challenge and MNIST data set. They also compare favorably against other state-of-the-art classifiers.

Detection and tracking of coronal mass ejections based on supervised segmentation and level set

Coronal mass ejection (CME) events refer to the appearance of a new, discrete, white-light feature (with outward velocity) in a coronagraph. The huge amount of data provided by the pertinent instruments onboard the Solar and Heliospheric Observatory (SOHO) and, most recently, the Solar Terrestrial Relations Observatory (STEREO) makes the human-based detection of such events excessively time consuming. Although several algorithms have been proposed to address this issue, there is still lack of universal consensus about their reliability. This work presents a novel method for the detection and tracking of CMEs as recorded by the LASCO instruments onboard SOHO. The algorithm we developed is based on level set and region competition methods, the CMEs texture being characterized by their co-occurrence matrix. The texture information is introduced in the region competition motion equations, and in order to evolve the curve, a fast level set implementation is used.

A decision step for Shape Context matching

This work presents a novel contribution in the field of shape recognition, in general, and in the Shape Context technique, in particular. We propose to address the problem of deciding if two shape context descriptors match or not using an a contrario approach. Its key advantage is to provide a measure of the quality of each match, which is a powerful tool for later recognition processes. We tested the proposed combination of Shape Context and the a contrario framework in character recognition from license plate images.

Skeleton-based action recognition using Citation-kNN on bags of time-stamped pose descriptors

With the advent of cost-effective depth sensors and the development of fast human-pose estimation algorithms, interest in action recognition from temporal skeleton sequences has been renewed. In this work we claim the task can be naturally seen as a Multiple Instance Learning (MIL) problem. Specifically, we model skeleton sequences as bags of time-stamped descriptors, and we present a new framework for action classification based on the Citation-kNN method. The proposed approach is effective in dealing with the large intra-class variability/inter-class similarity nature of the problem. Moreover, it is simple and provides a clear way for regulating tolerance to noise and temporal misalignment. Through extensive experiments on three datasets, we validate our approach and show that it compares favorably to other state-of-the-art skeleton-based action recognition methods.

Efficient descriptor tree growing for fast action recognition

Video and image classification based on Instance-to-Class (I2C) distance attracted many recent studies, due to the good generalization capabilities it provides for non-parametric classifiers. In this work we propose a method for action recognition. Our approach needs no intensive learning stage, and its classification performance is comparable to the state-of-the-art. A smart organization of training data allows the classifier to achieve reasonable computation times when working with large training databases. An efficient method for organizing training data in such a way is proposed. We perform thorough experiments on two popular action recognition datasets: the KTH dataset and the IXMAS dataset, and we study the influence of one of the key parameters of the method on classification performance.

Fast Non-parametric Action Recognition

In this work we propose a method for action recognition which needs no intensive learning stage, and achieves state-of-the-art classification performance. Our work is based on a method presented in the context of image classification. Unlike that method, our approach is well-suited for working with large real-world problems, thanks to an efficient organization of the training data. We show results on the KTH and IXMAS datasets. On the challenging IXMAS dataset, the average running time is reduced by 50% when using our method.

SAR Image Segmentation Using Level Sets and Region Competition under the

Synthetic Aperture Radar (SAR) images are dificult to segment due to their characteristic noise, called speckle , which is multiplicative, non-gaussian and has a low signal to noise ratio. In this work we use the

\mathcal{G}^H

\mathcal{G}^{H}

Model

distribution to model the SAR data from the different regions of the image. We estimate their statistical parameters and use them in a segmentation algorithm based on multiregion competition. We then apply this algorithm to segment simulated as well as real SAR images and evaluate the accuracy of the segmentation results obtained.