Hammadi Nait-Charif

Learning spatial context from tracking using penalised likelihoods

MAP estimation of Gaussian mixtures through maximisation of penalised likelihoods was used to learn models of spatial context. This enabled prior beliefs about the scale, orientation and elongation of semantic regions to be encoded, encouraging one-to-one correspondences between mixture components and these regions. In conjunction with minimum description length this enabled automatic learning of inactivity zones and entry zones from track data in a supportive home environment.

Automatic Estimation of Skeletal Motion from Optical Motion Capture Data

Utilization of motion capture techniques is becoming more popular in the pipeline of articulated character animation. Based upon captured motion data, defining accurate joint positions and joint orientations for the movement of a hierarchical human-like character without using a pre-defined skeleton is still a potential concern for motion capture studios. In this paper, we present a method for automatically estimating and determining the topology of hierarchical human skeleton from optical motion capture data based on the human biomechanical information. Through the use of a novel per-frame based recursive method with joint angle minimization, human skeleton mapping from optical marker and joint angle rotations are achieved in real time. The output of motion data from a hierarchical skeleton can be applied for further character motion editing and retargeting.

A Unified Framework for 3D Hand Tracking

Discriminative techniques are good for hand part detection, however they fail due to sensor noise and high inter-finger occlusion. Additionally, these techniques do not incorporate any kinematic or temporal constraints. Even though model-based descriptive for example Markov Random Field or generative for example Hidden Markov Model techniques utilize kinematic and temporal constraints well, they are computationally expensive and hardly recover from tracking failure. This paper presents a unified framework for 3D hand tracking, utilizing the best of both methodologies. Hand joints are detected using a regression forest, which uses an efficient voting technique for joint location prediction. The voting distributions are multimodal in nature; hence, rather than using the highest scoring mode of the voting distribution for each joint separately, we fit the five high scoring modes of each joint on a tree-structure Markovian model along with kinematic prior and temporal information. Experimentally, we observed that relying on discriminative technique i.e. joints detection produces better results. We therefore efficiently incorporate this observation in our framework by conditioning 50% low scoring joints modes with remaining high scoring joints mode. This strategy reduces the computational cost and produces good results for 3D hand tracking on RGB-D data.

Complex-Valued Neural Networks Fault Tolerance in Pattern Classification Applications

This paper investigates the fault-tolerance ability of complex-values neural networks (CVNNs) in classification applications. An analysis of the effect of weight loss at the units (neurons) level revealed that the loss of weight in complex neural networks is more critical than in real valued neural networks. A novel weight decay technique for fault tolerance of real-valued neural networks (RVNNs) is proposed and applied to CVNN. The simulation results indicate that the complex-valued neural networks are less fault tolerant than real-valued neural networks. It is also found that while the weight decay technique substantially improves the fault tolerance ability of RVNN, the technique does not necessary improve the fault tolerance of CVNNs.

Activity recognition for indoor fall detection using convolutional neural network

Falls are a major health problem in the elderly population. Therefore, a dedicated monitoring system is highly desirable to improve independent living. This paper presents a video based fall detection system in an indoor environment using convolution neural network. Identifying human poses is important in detecting fall events as specific “change of pose” defines a fall. Knowledge of series of poses is a key to detecting fall or non-fall events. A lying pose which may be considered as an after-fall pose is different from other normal activities such as lying/sleeping on the sofa or crawling. This paper uses Convolutional Neural Networks (CNN) to recognize different poses. Using Kinect, the following image combinations are explored: RGB, Depth, RGB-D and background subtracted RGB-D. We have constructed our own dataset by recording different activities performed by different people in different indoor set-ups. Our results suggest that combining RGB background subtracted and Depth with CNN gives the best possible solution for monitoring indoor video based falls.

Scatter reduction for grid-less mammography using the convolution-based image post-processing technique

X-ray Mammography examinations are highly affected by scattered radiation, as it degrades the quality of the image and complicates the diagnosis process. Anti-scatter grids are currently used in planar mammography examinations as the standard physical scattering reduction technique. This method has been found to be inefficient, as it increases the dose delivered to the patient, does not remove all the scattered radiation and increases the price of the equipment. Alternative scattering reduction methods, based on post-processing algorithms, are being investigated to substitute anti-scatter grids. Methods such as the convolution-based scatter estimation have lately become attractive as they are quicker and more flexible than pure Monte Carlo (MC) simulations. In this study we make use of this specific method, which is based on the premise that the scatter in the system is spatially diffuse, thus it can be approximated by a two-dimensional low-pass convolution filter of the primary image. This algorithm uses the narrow pencil beam method to obtain the scatter kernel used to convolve an image, acquired without anti-scatter grid. The results obtained show an image quality comparable, in the worst case, to the grid image, in terms of uniformity and contrast to noise ratio. Further improvement is expected when using clinically-representative phantoms.

Detailed Analysis of Scatter Contribution from Different Simulated Geometries of X-ray Detectors

Scattering is one of the main issues left in planar mammography examinations, as it degrades the quality of the image and complicates the diagnostic process. Although widely used, anti-scatter grids have been found to be inefficient, increasing the dose delivered, the equipment price and not eliminating all the scattered radiation. Alternative scattering reduction methods, based on post-processing algorithms using Monte Carlo MC simulations, are being developed to substitute anti-scatter grids. Idealized detectors are commonly used in the simulations for the purpose of simplification. In this study, the scatter distribution of three detector geometries is analyzed and compared: Case 1 makes use of idealized detector geometry, Case 2 uses a scintillator plate and Case 3 uses a more realistic detector simulation, based on the structure of an indirect mammography X-ray detector. This paper demonstrates that common configuration simplifications may introduce up to 14i¾?% of underestimation of the scatter in simulation results.