Cina Motamed

People counting via multiple views using a fast information fusion approach

Real-time estimates of a crowd size is a central task in civilian surveillance. In this paper we present a novel system counting people in a crowd scene with overlapping cameras. This system fuses all single view foreground information to localize each person present on the scene. The purpose of our fusion strategy is to use the foreground pixels of each single views to improve real-time objects association between each camera of the network. The foreground pixels are obtained by using an algorithm based on codebook. In this work, we aggregate the resulting silhouettes over cameras network, and compute a planar homography projection of each camera’s visual hull into ground plane. The visual hull is obtained by finding the convex hull of the foreground pixels. After the projection into the ground plane, we fuse the obtained polygons by using the geometric properties of the scene and on the quality of each camera detection. We also suggest a region-based approach tracking strategy which keeps track of people movements and of their identities along time, also enabling tolerance to occasional misdetections. This tracking strategy is implemented on the result of the views fusion and allows to estimate the crowd size dependently on each frame. Assessment of experiments using public datasets proposed for the evaluation of counting people system demonstrates the performance of our fusion approach. These results prove that the fusion strategy can run in real-time and is efficient for making data association. We also prove that the combination of our fusion approach and the proposed tracking improve the people counting.

Fast Moving Object Detection from Overlapping Cameras

In this work, we address the problem of moving object detection from overlapping cameras. We based on homographic transformation of the foreground information from multiple cameras to reference image. We introduce a new algorithm based on Codebook to get each single views foreground information. This method integrates a region based information into the original codebook algorithm and uses CIE L*a*b* color space information. Once the foreground pixels are detected in each view, we approximate their contours with polygons and project them into the ground plane (or into the reference plane). After this, we fuse polygons in order to obtain foreground area. This fusion is based on geometric properties of the scene and on the quality of each camera detection. Assessment of experiments using public datasets proposed for the evaluation of single camera object detection demonstrate the performance of our codebook based method for moving object detection in single view. Results using multi-camera open dataset also prove the efficiency of our multi-view detection approach.

A temporal fusion strategy for cross-camera data association

The purpose of this work is the design of a distributed vision system for vehicles tracking over wide areas. The tracking is performed by the re-identification of objects perceived by distant non-overlapping cameras. The data association is controlled by a temporal reasoning scheme. Decisions combine temporal and visual information. The visual information is composed by the 3D dimension and the normalised colour histogram of detected objects. Temporal constraints based on an acceleration model between nodes, are continuously updated with respect to the observed traffic behaviour. These constraints maintain a dynamic lifespan for all tracked objects. The management of the uncertainty represents an important component of the system. Statistical measurements are exploited at the sensor level information and a possibilistic approach permits to manage the ambiguities of the data association stage.

A multi-sensor validation approach for human activity monitoring

In this paper, we present a sensor validation approach for human activity monitoring. The sensor validation indicator is used for controlling the tracking process by a reactive strategy. The validation is represented by a confidence associated to each current tracked object and for each sensor and integrates static and dynamic contextual information. This confidence value is estimated by a Bayesian Network, which combines a set of heterogeneous information. The validation indicator takes into account the notion of geometrical visibility and the size of the detected object. We have tested this sensor validation approach in an indoor environment by using two complementary sensors for human activity monitoring. The first sensor is an industrial stereo camera (Triclops) and the second is a laser scanner (Sick LMS).

Adaptive Decision-level Fusion for Fongbe Phoneme Classification using Fuzzy Logic and Deep Belief Networks

In this paper, we compare three approaches for decision fusion in a phoneme classification problem. We especially deal with decision-level fusion from Naive Bayes and Learning Vector Quantization (LVQ) classifiers that were trained and tested by three speech analysis techniques: Mel-frequency Cepstral Coefficients (MFCC), Relative Spectral Transform - Perceptual Linear Prediction (Rasta-PLP) and Perceptual Linear Prediction (PLP). Optimal decision making is performed with the non-parametric and parametric methods. We investigated the performance of both decision methods with a third proposed approach using fuzzy logic. The work discusses the classification of an African language phoneme namely Fongbe language and all experiments were performed on its dataset. After classification and the decision fusion, the overall decision fusion performance is obtained on test data with the proposed approach using fuzzy logic whose classification accuracies are 95,54% for consonants and 83,97% for vowels despite the lower execution time of Deep Belief Networks.

Occlusion management in distributed multi-object tracking for visual-surveillance

This paper presents a distributed framework for multi-object tracking which deals with complex static and dynamic occlusions in visual-surveillance crowded scenes. Multiple autonomous particle filters are used for multi-object tracking in which each filter tracks a specific object. Stop-and-Go technique based on inter-blobs management, graph matching and a model of the scene is proposed for handling complex occlusions and inter-particle coalescence problems. The proposed technique is embedded into each autonomous filter to perform multi-object tracking in real time with linear complexity in terms of the number of the tracked objects. Experimental results in challengingsurveillance sequences demonstrate the robustness of the proposed framework.

A temporal belief-based hidden markov model for human action recognition in medical videos

In the context of human action recognition from video sequences in the medical environment, a Temporal Belief-based Hidden Markov Model (HMM) is presented. It allows to cope with human action temporality and enables to manage the data uncertainty and the knowledge incompleteness. The system of activity recognition is based on an HMM with explicit state duration. The global interpretation process uses the framework of the Transferable Belief Model (TBM). It enable us to model and manage the uncertainty over the video interpretation process. An application is proposed for human action analysis in medical video sequences provided by a patient monitoring system in the cardiology section in hospital. The proposed recognition method has been assessed on a database of 3000 video images of medical scenes and compared to the performance of the probabilistic Hidden Markov Models.

An Approach to Correcting Image Distortion by Self Calibration Stereoscopic Scene from Multiple Views

An important step in the analysis and interpretation of video scenes for recognizing scenario is the aberration corrections introduced during the image acquisition in order to provide and correct real image data. This paper presents an approach on distortion correction based on stereoscopic self-calibration from images sequences by using a multi-camera system of vision (network cameras). This approach for correcting image distortion brings an elegant and robust technique with good accuracy. Without any knowledge of shooting conditions, the cameras parameters will be estimated. For this, the image key points of interest are extracted from different overlapping views of multi-camera system by local descriptor, matching is realized between the images, and then the fundamental matrix is estimated and rectified if necessary. It is finally possible to calculate the cameras extrinsic and intrinsic parameters. These geometric information of the camera are used as parameters models of the distortion correction.

Automatic Scenario Recognition for Visual-Surveillance by Combining Probabilistic Graphical Approaches

The online recognition and indexing of video-surveillance sequence is firstly helpful for video-surveillance operator for an on-line alarm generation by highlighting abnormal situation. The second utility concerns the off-line retrieval of specific behavior from a stored image sequence in order to discover causes of an alarm. This capability becomes naturally more powerful when the monitoring concerns a network of IP-camera over a wide area or the Internet. The scenario recognition also known as activity recognition is an old and still active topic in computer science and several complementary approaches have been proposed by the Computer vision and the Artificial Intelligence communities. A scenario is composed on a set of elementary events linked with temporal constraints. The difficulty of human activity lies in their complexity, their spatial and temporal variability and also the uncertainty existing over the whole interpretation task. The computer vision approaches are generally focused on numerical approach by using probabilistic (Bui et al., 2004) (Hongeng et al., 2000) (Rabiner, 1989) or neural network (Howell & Buxton, 2002) approach in order to deal with uncertainty of the low level vision tasks. On the other hand, the Artificial Intelligence community has proposed more flexible symbolic approaches permitting a high level recognition capability (Tessier, 2003) (Vu et al., 2003) (Dousson & Maigat, 2007). Our main contribution in this work concerns the integration of these two complementary approaches (probabilistic and symbolic) in the global scenario recognition system. HHMs (Hidden Markov Model) are the most popular probabilistic approach in representing dynamic systems. They have been initially used in speech recognition (Rabiner, 1989) and successfully applied over gesture or activity recognition (Starner & Pentland, 1995). An interesting feature of HMM is its time scale invariance enabling activity with various speeds. Other extensions to the basic HMM have also been used such as the Coupled Hidden Markov Models (CHMMs) for modelling human interactions (Oliver et al., 2000), and variable length Markov models (VLMMs) to locally optimize the size of behavior models (Galata et al., 2001). However Bayesian networks have also been widely used in the computer vision community for object, event or scenario recognition. One important advantage of the Bayesian network is its ability to encode both qualitative and quantitative contextual knowledge, and their dependence.

Fuzzy-based algorithm for Fongbe continuous speech segmentation

Text-independent speech segmentation is a challenging topic in computer-based speech recognition systems. This paper proposes a novel time-domain algorithm based on fuzzy knowledge for continuous speech segmentation task via a nonlinear speech analysis. Short-term energy, zero-crossing rate and the singularity exponents are the time-domain features that we have calculated in each point of speech signal in order to exploit relevant information for generating the significant segments. This is down for the phoneme or syllable identification and the transition fronts. Fuzzy logic technique helped us to fuzzify the calculated features into three complementary sets namely: low, medium, high and to perform a matching phase using a set of fuzzy rules. The outputs of our proposed algorithm are silence, phonemes, or syllables. Once evaluated, our algorithm produced the best performances with efficient results on Fongbe language (an African tonal language spoken especially in Benin, Togo and Nigeria).