Although many elderly monitoring approaches either focus on activity or acoustic analysis, very limited work integrates both modalities within a unified framework for emergency detection. This paper presents a multi-modal method that combines skeleton-based activity with acoustic analysis to enhance the detection of medical emergencies in elderly individuals. A comprehensive pipeline is introduced, encompassing data preprocessing, feature extraction, multi-modal fusion, and emergency classification. Deep neural networks are leveraged to extract meaningful representations from both skeleton and acoustic data, while an advanced attention-based model refines the fusion process by prioritizing the most relevant features. The proposed method is validated through extensive experiments on specialized datasets for geriatric monitoring, demonstrating its effectiveness in accurately identifying abnormal activities and sounds associated with emergencies. Notably, the incorporation of attention mechanisms significantly improves detection reliability. This work contributes to the development of real-time continuous monitoring systems, promoting early intervention and reducing dependence on acute care.

Assisting visually impaired people to be more independent in their daily activities has always been a promising research area. In this context, we proposed a novel human action classification method to assist visually impaired people to understand the ongoing actions in their environment. The proposed method relies on the pretrained InceptionResNetV2 architecture to encode richer human action representations. The strength of this architecture relies on the combination of Inception and ResNet architectures to capture multi-scale features from a wide range of layers. Experimental evaluation was conducted on the Stanford40 actions dataset and proved the efficiency of the proposed method compared to the state-of-the-art human action classification methods in challenging contexts.

Gathering a dataset is essential in machine learning to develop new approaches and test their effectiveness. Despite their importance, it is difficult to find available datasets that meet researcher requirements, especially for the healthcare domain. This paper presents a new visual dataset of the top frequent elderly emergencies (heart attack, stroke, and Chronic Obstructive Pulmonary Disease (COPD) exacerbation). It is designed for the evaluation of different computational computer vision monitoring applications and human activity analysis. It can also be used for abnormal activity detection. This dataset covers the most frequent symptoms of each emergency. Each symptom is presented by different scenarios validated by a medical emergency expert. This dataset is recorded using the Kinect camera. Its samples are performed only in-depth and skeleton data to preserve the privacy of participants. Also, it is evaluated using methods based on Kinect skeleton data.

As life expectancy increases, people can live a longer life. However, ageing is associated with various health problems. As a result, older people are vulnerable to emergencies. To ensure the security of older people, it is necessary to analyse abnormal activities. Accordingly, vision sensors serve in elderly care and assisted living. Hence, this study allows the researchers to better understand the field of elderly emergency detection through abnormal activity analysis. In this work, we investigate and compare existing visual approaches and their performance. In addition, we explore available and interesting datasets. Moreover, we discuss existing approaches and datasets to improve real-world applications. We also present challenges to be tackled for future studies in monitoring applications.

Falls are considered one of the most severe health problems, especially among older people with physical disabili-ties. To ensure the security of the elderly, it is necessary to predict fall before it happens. This paper proposes a computer vision method for fall prediction among physically disabled elderly. Within our approach, we propose a novel implementation of Encoder-Decoder ConvLSTM (Convolutional Long Short-Term Memory). This work includes three parts: Data Acquisition, Data Preprocessing, and Data Analysis. Starting by acquiring skeleton streams using the Kinect camera. Then, applying preprocessing techniques: extracting skeleton features and selecting key frames. Finally, the analysis step consists of predicting the next frames and classifying them. In case of a predicted fall, an alert will be launched. To evaluate our approach, we use the FallFree dataset that covers all fall types and scenarios of people using canes and others without any mobility aid. Our method achieves an accuracy of 99.64%.

The moving object classification is a crucial step for several video surveillance applications whatever in the visible or thermal spectra. It still remains an active field of research considering the diversity of challenges related to this topic mainly in the context of an outdoor scene. In order to overcome several intricate situations, many moving objects classification methods have been proposed in the literature. Particular interest is given to the classes “Pedestrian” and “Vehicle”. In this paper, we have proposed a moving object classification approach based on deep learning methods from visible and infrared spectra. Three series of experiments carried on the challenging dataset “CD.net 2014” have proved that the proposed method reach accurate moving objects classification results when compared to methods based on deep learning and handcrafted features.

Human action recognition has received a lot of attention in computer vision community given its interest in many real applications. In this paper, we proposed a new method for human action recognition based on deep learning methods. The main contribution of the proposed method is an efficient combination of two Convolutional neural networks. The two-stream framework allows to fully utilize the rich multimodal information in videos. In fact, we explored the complementarity between appearance information and motion information to represent human actions. Specifically, we suggested a spatial Convolutional Neural Network performed on still individual images to model spatial information. To exploit motion between frames, a second Convolutional Neural Network is processed on accumulated optical flow images obtained by stacking the optical flow estimations between consecutive frames in a single image. Then, a fusion score is performed between the two Convolutional Neural Networks to achieve the appropriate class. In order to prove the performance of our method, we trained and evaluated our architecture on a standard human actions benchmark, the Weizmann dataset.

Intelligent video surveillance systems represent a potent tool for preserving human security in public places. Indeed, these surveillance systems are requested in several real-life scenarios, in order to assist security guards by alerting them in abnormal situations and helping them to retrieve a suspicious person. Especially, intelligent video surveillance systems based on UAV-sensors have the asset of monitoring large as well as difficult access spaces. In this scope, we introduce a new approach for suspicious person retrieval from UAV-sensors. The proposed approach implies two complementary phases which are an offline phase and an inference phase. Within these phases, a scene stabilization step is carried out. The offline phase allows building the non-person/person model as well as the person retrieval model. Nonetheless, the inference phase enables to detect persons and retrieve suspicious ones using the already generated models. The main contribution of the proposed approach is the use of part-level deep features in order to retrieve persons. The experimental results validate the contributions of our approach compared to the state-of-the-art approaches.

Person re-identification presents an active research area for intelligent video surveillance systems. The purpose is to find the same person from disjoint camera views at different times and locations. In this paper, we propose a novel Bi-Model Person Re-identification method (BiMPeR) that combines the appearance and the gait features to improve the re-identification performance and to handle the problem of similar appearances. The main idea is to prove the complementarity of these two modalities to extract a discriminative person signature for the re-identification problem. A score fusion method was adopted to combine these two modalities to reflect the impact of each one on the final decision. Experiments were performed on the CASIA-B database revealing promising results and showing the effectiveness of the proposed method against state-of-the-art uni-model methods.

People search based on semantic attributes presents an important task for several surveillance applications. The aim is to locate a suspect or to find a missing person in public areas based on an appearance description provided by a witness. In this paper, we propose a novel people search method based on an appearance description provided in terms of semantic attributes under uncontrolled acquisition conditions (e.g. gender, worn bags, carried objects, and clothes pattern). To this end, we propose to learn a set of deep semantic attribute classifiers based on the convolutional neural network. Experimental evaluations, based on the confusion matrix and the statistical tests, prove the efficiency of our method compared to state-of-the-art CNN architectures. Further, a comparison with state-of-the-art attribute classification methods is also conducted and confirms the efficiency of our method.

Summary The explosion of cloud services over the Internet has raised new challenges in cloud service selection and ranking. The existence of a great variety of offered cloud services made the users think deeply about the most appropriate services that meet their needs and at the same time are adaptable to their context. Nowadays, online reviews are used for the purpose of enhancing the effectiveness of finding useful product information, having impact on the consumers' decision‐making process. In this context, the current paper suggests a context‐aware cloud service ranking approach using online reviews and based on sentiment analysis (CROSA). Its main objective is to ease the cloud service selection. The CROSA approach analyzes sentiments associated with service measurement index (SMI)–based service properties for each alternative cloud service. Moreover, it enhances the cloud service decision‐making by supporting fuzzy sentiments through the intuitionistic fuzzy set theory and PROMETHEE II. The experimental results presented in this paper show that this approach is efficient and performing.

Gait has become a popular trait for biometric person recognition/re-identification. This is due to its advantage of being captured without any subject cooperation. This made it suitable especially for video surveillance applications. However, the gait features obtained in such scenarios depends on the observed walking direction of the subject. In this paper, we deal with the problem related to walking direction estimation in unconstrained environments. Covariates factors (i.e. carrying different types of bag, clothing) affect considerably the accuracy of walking direction estimation problem. Therefore, we have proposed a solution which is suitable for both real time application and unconstrained environment where the user walking direction is different and affected by covariates factors. The discriminative power of this solution is verified through experiments. The performance of this method was evaluated on the CASIA-B database. Experimental results prove the effectiveness of our proposed walking direction estimation method.

Nowadays, social media are taking an essential part of our daily life in several domains. Especially, on opinion analysis area with decisional systems where it helps companies to improve their products from reviews posted on social media such as Facebook, Twitter, etc. In this respect, there is an interest for opinion analysis or opinion mining expressed in social media for business decision support. In this research paper, we propose a new method of opinion analysis based on machine learning that determines the polarity of users'comments shared on different social media. The latter will be integrated in the ETL (Extract, Transform and Load) process to analyze the users' opinions. The proposed method is based on the n-grams technique to construct a semi-automatic dictionary for positive and negative keywords that is used in the learning phase to establish the prediction model. In addition, we propose a new features vector specific for social media for classifying the comments as positive, negative or neutral. The evaluation results performed on the both publicly data sets Stanford Twitter Sentiment (STS) and Sanders dataset showed a high accuracy level.

Learning from imbalanced data is attracting an increasing interest by the machine learning community. This is mainly due to the high number of real applications that are affected by this situation. The adaptation of the standard decision trees to deal with imbalanced data represents one of the important number of approaches that have been developed to address this problem. This adaptation has been proposed under three different perspectives: splitting criterion, assignment rule and pruning. In this paper, we focus our attention to the pruning of decision trees. We propose an adaptation of the standard pruning algorithm MCCP to address the skewed-data problem. Our contribution affects two levels: adaption of the metric used in selecting nodes to be firstly pruned and change of the evaluation measure used in selecting the best decision-tree through the pruning set. Our goal is to show that, contrary to the popular belief in the literature enquiring into the uselessness of decision tree pruning, an adaptive pruning technique for imbalanced situations is more efficient and more accurate towards the minority class. A total of twelve binary class data-sets having different imbalance ratio are used to test the performance of the proposed method. Experimental results show that the proposed post-pruning approach can increase the performance of imbalanced decision trees in terms of evaluation measures that are recent and appropriate for the context of imbalanced classification.

The increasing role of social media has pushed the scientific community to invest in them and take advantage of their information. Furthermore, the diffusion of social media has a very significant impact on several areas, such as marketing (customer's opinion analysis), terrorism, politics, etc. They affect a very wide audience, such as students, teenagers and companies. In fact, Facebook, Twitter and YouTube platforms play an increasingly important role for companie's to analysis customers generated content and also the information about their competitors. The transformation of data from social media into new knowledge for decision makers is a challenging task nowadays. In this context, several businesses have worked on these media to get extra information about their customers. Thus, the need for a data warehouse has become an emergency for the storage and management of data from social media. However, most of research studies have worked on only one social medium. Specifically, in this paper, we are interested in the design of multidimensional schema of data warehouse for opinion analysis. Firstly, we focuse on the design of the data mart from each social medium. Secondly, we present the data mapping between the multidimensional concepts of data mart determined from different social media. Thirdly, we are interested in the schema integration of several data mart schemes. Our goal is to achieve a generic data warehouse schema, including the most popular social media in order to analyze customers'opinions.

Person re-identification consists in recognizing the same person across non-overlapping camera views at different times and locations. It presents an important yet challenging task for intelligent video surveillance systems due to the large variations of pose, viewpoint, lighting and occlusion between the different camera views. Although a variety of algorithms have been presented in the past few years, most of them usually assume a pre-cropped bounding box of fully visible people taken from two different cameras to perform the re-identification process. However, in real-world video surveillance systems several challenges need to be addressed such as re-identifying truncated people and alleviating the hash lighting variation of the different camera views. In this paper, we propose an adaptive person re-identification approach that re-identifies a person irrespective of his status, i.e. truncated or fully visible, based on the apparent salient body regions driven by their robust appearance characteristics in a large camera network. The proposed approach has been experimentally validated on the High Definition Analytics (HDA) and viewpoint invariant perdestrian recognition data sets which include several re-identification challenges. The outcomes of this evaluation show promising results and demonstrate the effectiveness of our approach compared to other state-of-the-art approaches.

A biometric identification system is an automatic pattern recognition system that identifies a person through their specific physiological and/or behavioral characteristics. Unimodal biometric system often suffers from some limitations due to noise in sensed data, intra-class variation, inter-class similarities, spoof attacks, etc. Multi-biometric systems seek to overcome some of these limitations by providing multiple pieces of evidence about the person identity. The increased performance of multi-biometric systems motivated our investigation of a new approach for multi-biometric identification systems based on the combination of five finger surface instances of a person. The proposed approach has the particularity of using a new fusion scheme based on the rank level integration method to consolidate the results obtained from the different five biometric instances. To the best of our knowledge, this is the first time the rank level fusion approach is designed to incorporate the results of five finger surface instances to produce more reliable recognition results. To highlight the contributions of our approach, this paper presents a comparative experimental study of several rank level fusion approaches that can be useful in combining multi-biometric systems. The experimental evaluation was conducted on the real hand database ‘Sfax-Miracl hand database’. The experimental results suggest that our novel approach produces a significant performance improvement in the recognition accuracy over approaches based on individual finger surface. They also indicate that the rank level integration of the five finger surfaces is poised to provide a promising direction to finger surface-based multi-biometric systems.

We introduce a new computer vision based system for robust traffic sign recognition and tracking. Such a system presents a vital support for driver assistance in an intelligent automotive. Firstly, a color based segmentation method is applied to generate traffic sign candidate regions. Secondly, the HoG features are extracted to encode the detected traffic signs and then generating the feature vector. This vector is used as an input to an SVM classifier to identify the traffic sign class. Finally, a tracking method based on optical flow is performed to ensure a continuous capture of the recognized traffic sign while accelerating the execution time. Our method affords high precision rates under different challenging conditions.

In the recent years, the number of social media, such as Facebook, LinkedIn, Tumblr have been increasing regularly, have become ideal platforms for a large number of users at lower costs. However, social media helped companies get known by a wider public and collect the relevant information's about customers, their needs and their expectations. Therefore, several decision makers have worked on these media to get extra information about companies and customers. Moreover, the explosive growth of the amount of data makes their analysis and usage difficult. Therefore, the collection, analysis and the storage of these data in a data warehouse are extremely necessary. Thus, data warehouse modeling from social media should take into account aspects of social media. This paper proposes an approach for data warehouse construction from social media. Precisely, it deals with the conceptual modeling from social media, especially Facebook and Twitter. The proposed model can predict the success of the products prospects and improve the existing products.

This paper introduces a novel method of moving object classification in Infrared and Visible spectra. This method is based on a data-mining process by combining a set of best features based on shape, texture and motion. The proposed method relies either on visible spectrum or on infrared spectrum according to weather conditions (sunny days, rain, fog, snow, etc.) and timing of the video acquisition. Experimental studies are carried out to prove the efficiency of our predictive models to classify moving objects and the originality of our process with intelligent fusion of VIS-IR spectra.

Accidents caused by reduced concentration of drivers on traffic signs indications continue to represent an important part of accident-prone situations. Face to this threat, our work aims to develop a vision-based traffic sign recognition method based on a two-step recognition and 3D distance computing module. Firstly, a monocular color based segmentation method is applied to generate traffic sign candidates. Then, HoG features are applied to encode the detected traffic signs and compute the feature vector. This vector is used as an input to a SVM classifier to identify the traffic sign class. Secondly, a dense disparity map between the left and right images is created for the recognized traffic sign region to compute its distance to the vehicle carrying the stereovision. Our method affords high precision rates under different weather conditions. Moreover, it operates with a timing that is reasonable for real-time applications. The obtained results, compared to leading methods from the literature, prove the efficiency of our proposed method.

Hand detection is the first step of any hand biometric recognition process, which determines the outcome of the following treatments. In this paper, we propose a robust method for hand detection without contact and without constraints on the capture environment for hand biometric applications. This method is based on a color based approach adopting a non-parametric modeling. Our contribution consists mainly in the choice of the most relevant color axes and the choice of the decision rules automatically using a process of the data-mining as a new philosophy of data processing. To improve the achieved results of skin detection and to determine the hand region in the image, a succession of post-processing was proposed. Our hand detection method was evaluated experimentally on a real database; the outcomes of this evaluation show promising results and demonstrate the effectiveness of the proposed method.

Over the last two decades, the advances in computer vision and pattern recognition power have opened the door to new opportunity of automatic facial expression recognition system. In this work, we have introduced a new feature-based approach for facial expressions recognition. The proposed approach provides full automatic solution to identify human expressions as well as overcoming facial expressions variation and intensity problems. Facial features component were automatically detected and segmented. Then, we have detected facial feature points which go with facial expression deformations. Afterwards, distances between these points were computed and used through Data mining technique to generate a set of relevant prediction rules able to classify facial expressions. We took into account the intensity of JOY expression. Thus, we have defined SMILE expression as the lowest intensity of JOY. Seven facial expression classes were defined: JOY, SMILE, SURPRISE, DISGUST, ANGER, SADNESS, and FEAR We have appraised experimental study to evaluate the performance of the proposed solution.

This paper introduces a novel low-computation discriminative regions representation for expression analysis task. The proposed approach relies on interesting studies in psychology which show that most of the descriptive and responsible regions for facial expression are located around some face parts. The contributions of this work lie in the proposition of new approach which supports automatic facial expression recognition based on automatic regions selection. The regions selection step aims to select the descriptive regions responsible or facial expression and was performed using Mutual Information (MI) technique. For facial feature extraction, we have applied Local Binary Patterns Pattern (LBP) on Gradient image to encode salient micro-patterns of facial expressions. Experimental studies have shown that using discriminative regions provide better results than using the whole face regions whilst reducing features vector dimension.

Hand detection is the first step of any hand biometric recognition process, which determines the outcome of the following treatments. In this study, the authors propose a robust method for hand detection without contact and without constraints on the capture environment. This method is based on a data‐mining process for skin‐colour modelling. The presented data‐mining process offers several advantages like the choice of the most relevant colour axes and the automatic choice of the decision rules. To improve the achieved results of skin detection and to determine the hand region in the image, a succession of postprocessings was proposed. The authors hand detection method was evaluated experimentally on a real database, namely, ‘Sfax‐Miracl hand database’; the outcomes of this evaluation show promising results and demonstrate the effectiveness of the proposed method.

In this paper we deal with the problem of lowering down the difficulty of face detection in video. Most of the recently developed systems swap detection accuracy for higher speeds, or vice versa. We have proposed a robust approach which makes use of spatial and temporal information in video to reduce time execution and improve precision rate. Our experiments show that our proposed approach proves efficiency without sacrificing real-time performance which makes it well-suited for live video applications.

Obstacle detection is an important component in driver assistance as it helps systems to locate obstacles and then to prevent collisions. The aim of this study is to develop an obstacle detection module through digital images processing. We present a hybrid stereo vision-based method that combines stereo matching and homographic transformation methods. We use a sparse matching method in order to get a rapid geometric representation of the road scene that allows us to extract the upper and lower parts of obstacles. According to the position of the lower part, our method uses either the dense stereo matching or the homographic transformation methods to extract the candidate obstacles regions. A verification test is performed to verify whether the retained region is an obstacle or not. In order to avoid collisions, we compute the distance to the preceding obstacle to maintain the vehicle carrying the camera at a safety distance. The method presented here was tested on DIPLODOC road stereo sequence captured on a highway. The obtained results prove the efficiency of our proposed method.

The detection and extraction of characteristic features of the human face are primordial tasks in any given approach for face recognition. Within this context, the present paper will present a comprehensive list of steps that can help obtain the major discriminative features of the human face with the abstraction of the fuzzy data that are influenced by several external factors, notably light conditions, which often disrupt the results obtained by classifiers for human face recognition. The approach proposed in this paper can be considered a potentially strong candidate for use in a variety of commercial and industrial applications, particularly those related to security. In fact, in addition to its usefulness in identification processes, this face recognition system is also of particular importance for those who are interested in search and navigation processes in online video masses. In essence, our process was based on a corpus that contained a huge number of faces acquired in different positions and various lighting conditions. The identification and classification of faces was achieved through the use of a neural network called the multi-layer perceptron (MLP), which is one of the most common networks used in this context.

In this article, we present a contribution to the violent Web images classification. This subject is deeply important as it has a potential use for many applications such as violent Web sites filtering. We propose to combine the techniques of image analysis and data-mining to relate low level characteristics extracted from the image's colors to a higher characteristic of violence which could be contained in the image. We present a comparative study of different data mining techniques to classify violent Web images. Also, we discuss how the combination learning based methods can improve accuracy rate. Our results show that our approach can detect violent content effectively.

Along with the ever-growing Web comes the proliferation of objectionable content, such as sex, violence, racism, etc. We need efficient tools for classifying and filtering undesirable Web content. In this paper, we investigate this problem and describe WebGuard, an automatic machine learning-based pornographic Web site classification and filtering system. Unlike most commercial filtering products, which are mainly based on textual content-based analysis such as indicative keywords detection or manually collected black list checking, WebGuard relies on several major data mining techniques associated with textual, structural content-based analysis, and skin color related visual content-based analysis as well. Experiments conducted on a testbed of 400 Web sites including 200 adult sites and 200 nonpornographic ones showed WebGuard's filtering effectiveness, reaching a 97.4 percent classification accuracy rate when textual and structural content-based analysis was combined with visual content-based analysis. Further experiments on a black list of 12,311 adult Web sites manually collected and classified by the French Ministry of Education showed that WebGuard scored a 95.62 percent classification accuracy rate. The basic framework of WebGuard can apply to other categorization problems of Web sites which combine, as most of them do today, textual and visual content.

The paper presents a novel approach for robust skin-color detection using data-mining techniques. The goal of skin-color detection is to select the appropriate color model that allows pixels to be verified under different lighting conditions and other variations. When the appropriate color model is selected, it is implied that we have good skin-color classifier properties for skin detection. This model has been successfully applied to face detection and Web based adult content filtering issues.

@inproceedings{CI-CHAHIR-2003, author = {Hammami, M. and Chahir, Y. and Chen, L.}, title = {WebGuard: Web-Based Adult Content Detection and Filtering System}, booktitle = {IEEE/WIC International Conference on Web Intelligence (WI'03)}, pages = {574-578}, year = {2003}, address = {Halifax, Canada}, month = {October} }