Enrique Cabello

Recent advances in face biometrics with Gabor wavelets: A review

In this paper we focus on the great outburst of Gabor-based methods for face biometrics occurred in the last few years. Analytical approaches rely on the representation of a face with the Gabor responses computed on specific landmarks, while holistic methods take into account the face as a whole. We explore the role played by Gabor wavelets in international competitions, such as FERET or BANCA, where Gabor algorithms ranked first above other methods. By means of the analysis of five quantifiable factors, we present a ranking of methods as a function of their goodness. An enhanced version of AdaBoost, a complex-valued Gabor representation and a Gabor adaptive downsampling method are the three algorithms that lead the ranking. We also show there is a global trend toward face recognition methods, as well as toward Gabor holistic algorithms, due to their higher success rates.

HoGG: Gabor and HoG-based human detection for surveillance in non-controlled environments

A new method (HoGG) for human detection based on Gabor filters and Histograms of Oriented Gradients is presented in this paper. The effect of Gabor preprocessing is analyzed in detail, in particular the improvement experienced by the images information and the influence exerted over the extracted feature. To compare the performance of the proposed method, several alternative algorithms for human detection have been considered. In order to evaluate these techniques in non-controlled environments, a collection of standard databases, well known in the surveillance research community, has been used: PETS 2006, PETS 2007, PETS 2009 and CAVIAR. An exhaustive test design has been built based on two complementary evaluations: an evaluation oriented to counting people and a novel evaluation oriented to identification. Moreover, with the purpose of studying the performance of the Gabor-based preprocessing, a test adding Gabor filters to other local feature extraction methods, such as Steerable filters and the SIFT method, has been implemented. The HoGG method has achieved a good performance regardless of the difficulty of the images (occlusions, overlapping, carrying baggage, etc.). The proposed method has surpassed the alternative techniques in most of the analyzed situations. When the Gabor preprocessing is introduced into other local feature extraction methods, they achieve a better detection of the relevant information by enhancing the human shape. The results show that using Gabor preprocessing in techniques based on features like orientation or magnitude of gradient improve their performance. Given the excellent results obtained by HoGG at the identification-oriented evaluation, the method presented in this paper should be taken into account in the future design of intelligent surveillance systems.

Automatic 3d face feature points extraction with spin images

We present a novel 3D facial feature location method based on the Spin Images registration technique. Three feature points are localized: the nose tip and the inner corners of the right and left eye. The points are found directly in the 3D mesh, allowing a previous normalization before the depth map calculation. This method is applied after a preprocess stage where the candidate points are selected measuring curvatures on the surface and applying clustering techniques. The system is tested on a 3D Face Database called FRAV3D with 105 people and a widely variety of acquisition conditions in order to test the method in a non-controlled environment. The success location rate is 99.5% in the case of the nose tip and 98% in the case of eyes, in frontal conditions. This rate is similar even if the conditions change allowing small rotations. Results in more extremely acquisition conditions are shown too. A complete study of the influence of the mesh resolution over the spin images quality and therefore over the face feature location rate is presented. The causes of the errors are discussed in detail.

Influence of wavelet frequency and orientation in an SVM-based parallel Gabor PCA face verification system

We present a face verification system using Parallel Gabor Principal Component Analysis (PGPCA) and fusion of Support Vector Machines (SVM) scores. The algorithm has been tested on two databases: XM2VTS (frontal images with frontal or lateral illumination) and FRAV2D (frontal images with diffuse or zenithal illumination, varying poses and occlusions). Our method outperforms others when fewer PCA coefficients are kept. It also has the lowest equal error rate (EER) in experiments using frontal images with occlusions. We have also studied the influence of wavelet frequency and orientation on the EER in a one-Gabor PCA. The high frequency wavelets are able to extract more discriminant information compared to the low frequency wavelets. Moreover, as a general rule, oblique wavelets produce a lower EER compared to horizontal or vertical wavelets. Results also suggest that the optimal wavelet orientation coincides with the illumination gradient.

Detecting denial of service by modelling web-server behaviour

The article presented here discusses a system which characterizes HTTP traffic and discriminates between legitimate and other kinds of HTTP traffic, such as those generated by Botnets or distributed denial of service (DDoS) tools. The system presented in this paper uses three analyses that are sequentially applied to the traffic flow to detect abnormal users. Combining statistical methods as well as analysis of HTTP request paths and the access time to the different resources in the web server, we have labelled abnormal users in real traffic flow. First, we have tested our prototype in real traffic from a multi-site web server detecting all abnormal users, such as an illegitimate audit of the server, Google bot and web-crawlers. In a second experiment, the most common DDoS attacks were introduced in the real traffic flow. As a result, all suspicious users were detected and labelled.

Multimodal 2D, 2.5D & 3D Face Verification

A multimodal face verification process is presented for standard 2D color images, 2.5D range images and 3D meshes. A normalization in orientation and position is essential for 2.5D and 3D images to obtain a corrected frontal image. This is achieved using the spin images of the nose tip and both eyes, which feed an SVM classifier. First, a traditional principal component analysis followed by an SVM classifier are applied to both 2D and 2.5D images. Second, an iterative closest point algorithm is used to match 3D meshes. In all cases, the equal error rate is computed for different kinds of images in the training and test phases. In general, 2.5D range images show the best results (0.1% EER for frontal images). A special improvement in success rate for turned faces has been obtained for normalized 2.5D and 3D images compared to standard 2D images.

Driving risk classification based on experts evaluation

A novel multidisciplinary system for the automatic driving risk level classification is presented. The data considered involves the three basic traffic safety elements (driver, road, and vehicle), as well as knowledge from traffic experts. The driving experiments were conducted in a truck cabin simulator handled by a professional driver, considering the most common real-world enviroments. Each traffic expert evaluate the driving risk on a 0 to 100 visual analogue scale. The driver, road and vehicle information was used to train five different data mining algorithms in order to predict the driving risk level. The benefits of the completeness of the data considered in our system are presented and discussed.

A New Approach to Identify Big Rocks with Applications to the Mining Industry

Detection of big rocks is an important, even critical, problem in the mining industry due to the risk of machine blockage causing high costs. This paper presents a computer-vision-based method to detect big rocks in a real mining industry. Our system, based on a mixture of image processing techniques and neural networks, works as follows: once the image is taken, a pre-processing step is performed, filtering the image and extracting a set of candidate rocks. Then a neural network processes the candidate rocks to ensure correct detection. A tracking algorithm is then applied to avoid false detection due to rock grouping. Using geometrical information, it is possible to estimate the real dimensions of the rocks. Our computer vision system satisfies time constraints imposed by the industry to work in real time and is currently operating. The algorithm presented is independent of the rocks shape. Results obtained during nine months of unsupervized work are provided, showing that our system is able to work under different light conditions and is robust enough to face real work conditions.

PCA vs low resolution images in face verification

Principal components analysis (PCA) has been one of the most applied methods for face verification using only 2D information, in fact, PCA is practically the method of choice for face verification applications in the real-world. An alternative method to reduce the problem dimension is working with low resolution images. In our experiments, three classifiers have been considered to compare the results achieved using PCA versus the results obtained using low resolution images. An initial set of located faces has been used for PCA matrix computation and for training all classifiers. The images belonging to the testing set were chosen to be different from the training ones. Classifiers considered are k-nearest neighbours (KNN), radial basis function (RBF) artificial neural networks, and support vector machine (SVM). Results show that SVM always achieves better results than the other classifiers. With SVM, correct verification difference between PCA and low resolution processing is only 0.13% (99.52% against 99.39%).

Person detection in surveillance environment with HoGG: Gabor filters and Histogram of Oriented Gradient

In this paper, a novel method (HoGG) for human detection in non-controlled surveillance environments based on Gabor filters and Histograms of Oriented Gradients is proposed. Complete performance results of several alternative algorithms for human detection are presented. In order to evaluate these methods in non-controlled situations, a set of standard video surveillance databases (PETS 2006, PETS 2007, PETS 2009 and CAVIAR) as well as the classical (in people detection area) INRIA database are used. An exhaustive test design has been built based on two complementary evaluations: an evaluation oriented to count persons and a novel evaluation oriented to identification. The proposed method has surpassed the alternative methods in most of the analyzed situations, this is a relevant improvement of the widely used HoG algorithm.