Johel Miteran

Definition and Performance Evaluation of a Robust SVM Based Fall Detection Solution

We propose an automatic approach to detect falls in home environment. A Support Vector Machine based classifier is fed by a set of selected features extracted from human body silhouette tracking. The classifier is followed by filtering operations taking into account the temporal nature of a video. The features are based on height and width of human body bounding box, the users trajectory with her/his orientation, Projection Histograms and moments of order 0, 1 and 2. We study several combinations of usual transformations of the features (Fourier Transform, Wavelet transform, first and second derivatives), and we show experimentally that it is possible to achieve high performance using a single camera.We evaluated the robustness of our method using a realistic dataset. Experiments show that the best tradeoff between classification performance and time processing result is obtained combining the original data with their first derivative. The global error rate is lower than 1%, and the recall, specificity and precision are high (respectively 0.98, 0.996 and 0.942). The resulting system can therefore be used in a real environment. Hence, we also evaluated the robustness of our system regarding location changes. We proposed a realistic and pragmatic protocol which enables performance to be improved by updating the training in the current location, with normal activities records.

Optimized spatio-temporal descriptors for real-time fall detection: comparison of support vector machine and Adaboost-based classification

Abstract. We propose a supervised approach to detect falls in a home environment using an optimized descriptor adapted to real-time tasks. We introduce a realistic dataset of 222 videos, a new metric allowing evaluation of fall detection performance in a video stream, and an automatically optimized set of spatio-temporal descriptors which fed a supervised classifier. We build the initial spatio-temporal descriptor named STHF using several combinations of transformations of geometrical features (height and width of human body bounding box, the user’s trajectory with her/his orientation, projection histograms, and moments of orders 0, 1, and 2). We study the combinations of usual transformations of the features (Fourier transform, wavelet transform, first and second derivatives), and we show experimentally that it is possible to achieve high performance using support vector machine and Adaboost classifiers. Automatic feature selection allows to show that the best tradeoff between classification performance and processing time is obtained by combining the original low-level features with their first derivative. Hence, we evaluate the robustness of the fall detection regarding location changes. We propose a realistic and pragmatic protocol that enables performance to be improved by updating the training in the current location with normal activities records.

Classification of prostate magnetic resonance spectra using support vector machine

Prostate cancer is the most common cancer in men over 50 years of age and it has been shown that nuclear magnetic resonance spectra are sensitive enough to distinguish normal and cancer tissues. In this paper, we propose a classification technique of spectra from magnetic resonance spectroscopy. We studied automatic classification with and without quantification of metabolite signals. The dataset is composed of 22 patient datasets with a biopsy-proven cancer, from which we extracted 2464 spectra from the whole prostate and of which 1062 were localised in the peripheral zone. The spectra were manually classed into 3 different categories by a spectroscopist with 4 years experience in clinical spectroscopy of prostate cancer: undetermined, healthy and pathologic. We used different preprocessing methods (module, phase correction only, phase correction and baseline correction) as input for Support Vector Machine and for Multilayer Perceptron, and we compared the results with those from the expert. If we class only healthy and pathologic spectra we reach a total error rate of 4.51%. However, if we class all spectra (undetermined, healthy and pathologic) the total error rate rises to 11.49%. We have shown in this paper that the best results are obtained using the pre-processed spectra without quantification as input for the classifiers and we confirm that Support Vector Machine are more efficient than Multilayer Perceptron in processing high dimensional data.

Automatic hardware implementation tool for a discrete Adaboost-based decision algorithm

We propose a method and a tool for automatic generation of hardware implementation of a decision rule based on the Adaboost algorithm. We review the principles of the classification method and we evaluate its hardware implementation cost in terms of FPGAs slice, using different weak classifiers based on the general concept of hyperrectangle. The main novelty of our approach is that the tool allows the user to find automatically an appropriate tradeoff between classification performances and hardware implementation cost, and that the generated architecture is optimized for each training process. We present results obtained using Gaussian distributions and examples from UCI databases. Finally, we present an example of industrial application of real-time textured image segmentation.

Fast prototyping of a SoC-based smart-camera: a real-time fall detection case study

Smart camera, i.e. cameras that are able to acquire and process images in real-time, is a typical example of the new embedded computer vision systems. A key example of application is automatic fall detection, which can be useful for helping elderly people in daily life. In this paper, we propose a methodology for development and fast-prototyping of a fall detection system based on such a smart camera, which allows to reduce the development time compared to standard approaches. Founded on a supervised classification approach, we propose a HW/SW implementation to detect falls in a home environment using a single camera and an optimized descriptor adapted to real-time tasks. This heterogeneous implementation is based on Xilinx’s system-on-chip named Zynq. The main contributions of this work are (i) the proposal of a co-design methodology. These methodologies enable the HW/SW partitioning to be delayed using high-level algorithmic description and high-level synthesis tools. Our approach enables fast prototyping which allows fast architecture exploration and optimisation to be performed, (ii) the design of a hardware accelerator dedicated to boosting-based classification, which is a very popular and efficient algorithm used in image analysis, (iii) the proposal of fall-detection embedded in a smart camera and enabling integration into the elderly people environment. Performances of our system are finally compared to the state-of-the-art.

Detection and matching of curvilinear structures

We propose an approach to curvilinear and wiry object detection and matching based on a new curvilinear region detector (CRD) and a shape context-like descriptor (COH). Standard methods for local patch detection and description are not directly applicable to wiry objects and curvilinear structures, such as roads, railroads and rivers in satellite and aerial images, vessels and veins in medical images, cables, poles and fences in urban scenes, stems and tree branches in natural images, since they assume the object is compact, i.e. that most elliptical patches around features cover only the object. However, wiry objects often have no flat parts and most neighborhoods include both foreground and background. The detection process is first evaluated in terms of segmentation quality of curvilinear regions. The repeatability of the detection is then assessed using the protocol introduced in Mikolajczyk et al. [1]. Experiments show that the CRD is at least as robust as to several image acquisition conditions changes (viewpoint, scale, illumination, compression, blur) as the commonly used affine-covariant detectors. The paper also introduces an image collection containing wiry objects and curvilinear structures (the W-CS dataset).

An FPGA-based accelerator for Fourier Descriptors computing for color object recognition using SVM

Fourier Descriptors (FD) can be used as feature vector components in various applications, such as real-time color object recognition or image retrieval. The full process is composed of the feature extraction followed by a classification step performed using support vector machine (SVM). In order to accelerate the computation of FD, a hardware implementation using FPGA technology is presented in this paper. We evaluated classification performance with respect to lighting variations and noise sensibility. Several experiments were carried out on three databases. Then an efficient architecture for FD computation on FPGAs is proposed and designed as accelerator. The WildCard is used to prototype this implementation. This design can have an operation speed up of approximately 10 compared to the standard software PC implementation.

Configurable motion-estimation hardware accelerator module for the MPEG-4 reference hardware description platform

This paper describes a motion estimation co-processor architecture that explicitly separates the implementation stages consisting of data access to the search window and the evaluation of the matching criterion from the implementation of the search strategy. The architecture is modular and can be re-configured according to the different MPEG video coding profiles and level parameters. Although the architectural solutions described here can be in principle applied to any SoC implementation technologies, the solution presented here is expressly conceived and validated on FPGA co-processing architectures supporting mixed SW/HW implementations of video encoders such as generic PC platforms with a standard PCMCIA FPGA cards. The module has been developed in the framework of the MPEG reference hardware description activity.

SVM approximation for real-time image segmentation by using an improved hyperrectangles-based method

A real-time implementation of an approximation of the support vector machine (SVM) decision rule is proposed. This method is based on an improvement of a supervised classification method using hyperrectangles, which is useful for real-time image segmentation. The final decision combines the accuracy of the SVM learning algorithm and the speed of a hyperrectangles-based method. We review the principles of the classification methods and we evaluate the hardware implementation cost of each method. We present the combination algorithm, which consists of rejecting ambiguities in the learning set using SVM decision, before using the learning step of the hyperrectangles-based method. We present results obtained using Gaussian distribution and give an example of image segmentation from an industrial inspection problem. The results are evaluated regarding hardware cost as well as classification performances.

Access control: adaptation and real-time implantation of a face recognition method

An improvement of a face recognition method is proposed. The goal is to develop easy-to-use access control software, allowing personal-computer or building access control with minimal constraints for the users. This approach requires a high-speed classification method (about 8 images/s) and a high global recognition rate. We obtain good results using a method derived from principal-component analysis, a geometric transformation of the feature space, and a fast decision algo- rithm.