Anca Apatean

Combining SURF-based local and global features for road obstacle recognition in far infrared images

This paper describes a road obstacle classification system that recognizes both vehicles and pedestrians in far-infrared images. Different local and global features based on Speeded Up Robust Features (SURF) were investigated and then selected in order to extract a discriminative signature from the infrared spectrum. First, local features representing the local appearance of an obstacle, are extracted from a codebook of scale and rotation-invariant SURF features. Second, global features were used since they provide complementary information by characterizing shape and texture. When compared with the state-of-the-art Haar and Gabor wavelet features, our method provides significant improvement of recognition performances. Moreover, since our SURF based representation is invariant to the scale and the number of local features extracted from objects, our system performs the recognition task without resizing images. Our system was evaluated on a set of far-infrared images where obstacles occur at different scales and in difficult recognition situations. By using a multi-class SVM approach, accuracy rates of 91.51% has been achieved on Surf-based representation, while a maximum rate of 89.11% was achieved on wavelet-based representation.

Visible-infrared fusion in the frame of an obstacle recognition system

In this article we propose different fusion schemes using information provided by visible and infrared images for road obstacle SVM-based classification. Three approaches for the fusion of VIS and IR information are presented. The early fusion yields a feature vector integrating at the feature level both visual and infrared information. The obtained bimodal feature vector is used as input to an SVM-based classification scheme. The intermediate fusion, which is performed at the kernel level combines different simple kernels of the SVM classifier in order to obtain a multiple kernel (MK). The late fusion combines matching scores of individual obstacle recognition modules in order to improve the systems final decision. In this late fusion case two methods have been considered to calculate the optimum weighting parameter: an Adaptive Fusion of Scores (AFScores) and a non-Adaptive Fusion of Scores (nAFScores). Comparative results showed that fusion-based obstacle recognition systems outperform monomodal visual and infrared obstacle recognizers. An important advantage of these fusion-based systems is their possibility to adapt to the environmental illumination conditions due to the weighting parameter which can contribute to the adjustments of the systems final decision.

Objects recognition in visible and infrared images from the road scene

The detection of an obstacle in a traffic scene situation (obstacle which most often means a pedestrian or a vehicle) is a real challenge due to the outdoor environment and the variety of appearance of the obstacle. In this paper some details about our recognition module applied on visible and infrared image databases are presented. Given an image, or a region within an image, generate different types of features (Haar and Gabor wavelet, seven statistics moments, eight most important DCT coefficients and some GLCM coefficients) that will be fed to a classifier, in order to classify the image in one of the 5 possible classes: standing person, unknown posture, motor bike, tourism car and utility car. Different types of classifiers (KNN and SVM with an RBF kernel) were used to examine the data. Accuracy rates above 92% have been achieved.

Bimodal approach in emotion recognition using speech and facial expressions

This paper aims to present a multimodal approach in emotion recognition which integrates information from both facial expressions and speech signal. Using two acted databases on different subjects, we were able to emphasize six emotions: sadness, anger, happiness, disgust, fear and neutral state. The models in the system were designed and tested by using a Support Vector Machine classifier. Firstly, the analysis of the strengths and the limitations of the systems based only on facial expressions or speech signal was performed. Data was then fused at the feature level. The results show that in this case the performance and the robustness of the emotion recognition system have been improved.

Obstacle recognition using multiple kernel in visible and infrared images

We propose a fusion model at data-level based on a linear combination of kernels for an SVM-based classification. The kernel functions are evaluated on disjoint entries, on the signature acquired from the visible and infrared spectrum. Different feature extraction and feature selection algorithms have been investigated in order to compute different feature vectors. A bi-objective optimization (using accuracy rate and classification time) is used to assure the kernel selection, the hyperparameters optimization but also the adaptation of the system to different difficult conditions using the sensor weighting coefficient. Our purpose is to develop the obstacle recognition module and to obtain a robust model for an SVM-multiple-kernel based classification.

Information fusion for obstacle recognition in visible and infrared images

We propose the information fusion of visible and infrared images for a pedestrian-vehicle SVM-based classification. Different types of fusion methods are presented: data fusion, feature fusion, matching score fusion and decision fusion. Data - level fusion assumes that the raw information is combined at the pixel level. The fusion at the feature level produces a feature vector integrating both visual and infrared information. Matching score fusion and decision fusion combine matching scores or decisions of individual obstacle recognition modules. Comparative results showed that fusion-based obstacle recognition techniques outperformed individual visual and infrared obstacle recognizers. An important advantage of these fusion-based systems is their possibility to adapt to the environmental illumination conditions due to a weighting parameter which also controls the systems final decision. Different feature extraction and feature selection algorithms have been investigated in order to retain the best suited features for the classification process.

Speaker diarization experiments for Romanian parliamentary speech

The speaker diarization system LIUM was developed for broadcast news data, but in this paper we employed it to evaluate the diarization performance over Romanian audio files of parliamentary speeches with different challenges (interruptions, noise, overlapping of speaker discourses etc.). More than 21 hours of speech belonging to few hundreds of speakers were used to evaluate the automatic process of segments assignation. The obtained results demonstrated that LIUM is proper even for parliamentary speeches with multiple challenges.

Wavelets and moments for obstacle classification

The artificial vision systems was developed having as model the human system, and therefore the objects recognition task is reduced to a classification: the recognition of an initial unknown object through detection of the similarities to another object, previously learned. Our purpose is to study the obstacle recognition in the ruttier scene using wavelet transform. We compared different recognition rates obtained by the use of different mother wavelet functions (as Daubechies, Coiflet, Biorthogonal and the recent discovered ones, named fractional B- splines). In order to improve the recognition rates, we added first order statistics features and the seven moments of Hu.

Kernel and Feature Selection for Visible and Infrared based Obstacle Recognition

In this article we propose a fusion model at data-level based on a linear combination of kernels. These kernels functions will be evaluated on disjoint entries, on the signature acquired from visible respective infrared spectrum. Therefore, we have to choose the proper numeric signature for the visible and for the infrared images. In order to retain just the best suited features, different feature extraction and feature selection algorithms have been investigated. In this way, important information can be achieved in a small number of coefficients, implying thus a significant reduction of the computation time. Our purpose is to develop the obstacle recognition module and to examine if a visible-infrared fusion is efficient for this task.

A Model to Measure the Performance of Human Resources in Organisations

Abstract The economic crisis, demography, technology, globalization etc. are all factors which will influence the organizational structures and business strategies. A new business strategy will require, among others, that passive Human Resources Management (HRM) change into an active one with a decisive influence upon business. The vision of an active HRM requires that HR information (IT) dedicated systems assist human resources managers in their decision-making. The existing IT systems predominantly manage the salary calculations and, possibly, the employees professional development, two of the tasks that a human resources manager has to pursue. However, tasks such as assisting, consulting and engaging the human resources in the organization are equally important. IT systems must also develop into these directions. The present paper proposes a solution to measure the performance of human resources by creating an employee performance indicator (EPI). The paper first describes the economic phenomenon involved in the HR performance process, then the mathematical model is formulated, the algorithm is implemented, the solution of the model is analysed from a technical and economic point of view, and finally the decision is made. We use the weighted arithmetic mean to compute the EPI indicator and the correlation formula to establish the degree of relevance between the EPI indicator and the variables involved in the model. An implementation in R is given.