Carlos A. Holzmann

Face and eye tracking algorithm based on digital image processing

A non-invasive interface to track eye position using digital image processing techniques is under development. Information about head and eye position is obtained from digital images. The objective is to develop an interface to detect eye position based only on digital image processing algorithms, free of electrodes or other electronic devices. We propose a method for eye tracking built into five stages. These include: coarse and fine face detection, finding the eye region of maximum probability, map of the pupil/iris location and pupil/iris detection. Using frontal face images obtained from a database, the probability maps for the eye region were built. Only gray levels are considered for this computation (8 bits). The algorithms for face and eye detection were assessed on 102 images from the Purdue database and on 897 images from a video sequence. The face detection algorithm reached a 99% and 100% correct detection rate on the databases respectively. On the same databases the pupil/iris detection algorithm reached 85.3% and 98.4% of correct detection respectively.

Expert-system classification of sleep/waking states in infants

This work is part of a project to develop an expert system for automated classification of the sleep/waking states in human infants; i.e. active or rapid-eye-movement sleep (REM), quiet or non-REM sleep (NREM), including its four stages, indeterminate sleep (IS) and wakefulness (WA). A model to identify these states, introducing an objective formalisation in terms of the state variables characterising the recorded patterns, is presented. The following digitally recorded physiological events are taken into account to classify the sleep/waking states: predominant background activity and the existence of sleep spindles in the electro-encephalogram; existence of rapid eye movements in the electro-oculogram; and chin muscle tone in the electromyogram. Methods to detect several of these parameters are described. An expert system based on artificial ganglionar lattices is used to classify the sleep/waking states, on an off-line minute-by-minute basis. Algorithms to detect patterns automatically and an expert system to recognise sleep/waking states are introduced, and several adjustments and tests using various real patients are carried out. Results show an overall performance of 96.4% agreement with the expert on validation data without artefacts, and 84.9% agreement on validation data with artefacts. Moreover, results show a significant improvement in the classification agreement due to the application of the expert system, and a discussion is carried out to justify the difficulties of matching the experts criteria for the interpretation of characterising patterns.

Two-point vibrotactile discrimination related to parameters of pulse burst stimulus

Tactile spatial resolution is an important factor in the design of vibrotactile arrays. The two-point discrimination distance is used as a measure of tactile spatial resolution. An experimental study is presented showing the effect of pulse burst stimulus parameters, pulse repetition period and duty cycle on two-point vibrotactile spatial discrimination. An array of piezoceramic vibrators is used to measure two-point spatial discrimination on the index finger. In a group of 14 subjects, the average two-point discrimination distance for a pulse repetition period of 1/25 s is 2.1 mm (SD=1.0), whereas for 1/500 s it is 5.1 mm (SD=0.9). Differences in discrimination distances are statistically significant according to the ANOVA analysis (p<0.001). Results show that the two-point discrimination distance is better for longer pulse repetition periods. Therefore the pulse repetition period in an excitatory waveform composed of bursts of pulses is important for tactile resolution. No statistically significant differences in discrimination distances are found between bursts of pulses of 50% duty cycle and those of lower duty cycle. The latter result indicates that, by choosing low-duty cycle waveforms for vibrotactile stimulation, the power can be reduced with no loss in two-point discrimination capacity.

Improvements on handwritten digit recognition by genetic selection of neural network topology and by augmented training

The paper presents two ways of improving the handwritten digit recognition ability of a neural network. First by selection of the number of hidden units in the neural network, and second, by training using an augmented set of patterns. The handwritten digit recognition application is performed on feed-forward, fully connected neural networks with two hidden layer architectures. A genetic algorithm is used to search among configurations of two unequal hidden layer networks to find the optimum number of hidden units. Training procedures involving augmented sets of training patterns are produced by two methods: by shifting and by magnification every handwritten digit of the original training set. Results show that the best two hidden layer network, 178/spl times/26 units, trained for 10 different random starting weight sets in centered mode, i.e., no shifting, resulted in 82.74% average recognition rate with standard deviation, STD=0.66. The best two hidden layer network, 154/spl times/58, trained for 11 different random starting weight sets in shifting mode, resulted in 92.09% average recognition rate with STD=0.37. A comparison of recognition rates for centered versus shifting training modes for the genetic selection resulted statistically significant with p<0.001.

Classification of sleep stages in infants: a neuro fuzzy approach

An ANFIS based neuro-fuzzy system to classify sleep-waking states and stages in healthy infants has been developed. The classifier takes rive input patterns identified from polysomnographic recordings on 20 s frames and assigns them to one out of rive possible classes (WA, NREM-I, NREM-II, NREM-III&IV or REM). Eight polysomnographic recordings of healthy infants were studied, making a total of 3510 frames. Of these, four recordings were used for training, two for validation and two for testing. Results on the testing data achieved on average 88.2% of expert agreement in sleep-waking state-stage classification. These results were compared with the ones obtained using a multi-layer perceptron neural network (87.3%) and by applying the experts rules for sleep classification (86.7%). The neuro-fuzzy approach also rendered fuzzy classification rules, which were analyzed and compared with the experts rules.

Design of a virtual keyboard based on iris tracking

The development of man-machine interfaces to control devices with the eyes could be of great impact in handicapped individuals. In this paper a non-obstructive interface is proposed to detect and track iris position based on digital image processing techniques, templates and a reference point mounted near the eye. The position of the iris is detected in four steps: reference detection, iris center detection, iris position computation relative to the reference and determination of the eye position within the virtual keyboard. The position of the iris is projected over a virtual keyboard and it is determined the maximum number of keys that the system is able to discriminate. The percentage of iris correct detection in four different video sequences of 803, 710, 913, 849 frames, respectively were above 98.2%. The reference detection was 100% correct. A virtual keyboard was built allowing 12 horizontal and 8 vertical keys.

Improvements on handwritten digit recognition by cooperation of modular neural networks

In this paper modular neural networks are used to improve handwritten digit recognition. To evaluate the performance of modular networks, a comparison is made with a global neural network, on the same database. Two basic kind of modular networks are considered: 1) seven expert modular networks in which five of them are provided for digits 0, 1, 2, 5, 6, 7 and the rest for the pair of digits 3-8 and 4-9 respectively; and 2) a modular neural network with an expert module for each feature extracted from the handwritten digit image. The cooperation is among modules extracting slope and radial projection from each digit. Two type of cooperation among modular networks are considered: neural network and weighted combination of the modules outputs. The models were trained and tested on a different set of digits. The results show that by using modular network for features, it is possible to improve classification performance on handwritten digits, from 91.0% in the case of global networks to 93.5% of modular networks.

Power requirements for vibrotactile piezo-electric and electromechanical transducers.

Human-machine information transfer through tactile excitation has addressed new applications in virtual reality, robotics, telesurgery, sensory substitution and rehabilitation for the handicapped in the past few years. Power consumption is an important factor in the design of vibrotactile displays, because it affects energy needs and the size, weight, heat dissipation and cost of the associated electronics. An experimental study is presented on the power required to reach tactile thresholds in electromechanical and piezo-electric transducers. Three different waveforms are considered, with an excitatory period formed by a burst of rectangular 50% duty cycle pulses (R50), rectangular low duty cycle pulses (RLO) and sinusoidal pulses (SIN). Ten different pulse repetition periods (RPs) were considered in the range 1/550-1/25s. The voltage and current waveforms applied to the transducers at sensation thresholds in a group of 12 healthy subjects were sampled and stored in a digital oscilloscope. The average power was determined for each subject, and differences of two orders of magnitude were measured between the electromechanical and the piezo-electric transducer power consumption. Results show that, for the electromechanical transducer, a smaller power consumption of 25μW was determined for RP=1/25s and the RLO waveform. In the case of the piezo-electric transducer, power of 0.21 μW was determined for SIN excitation and RP=1/250s. These results show the advantages of reducing power requirements for vibrotactile displays, which can be optimised by the choice of appropriate types of transducer, excitatory waveforms and pulse repetition periods.

Optimization of the envelope of a short duty cycle pulse waveform for tactile stimulation

Presents a study to optimize the power delivered by an excitatory waveform to the frequency region of maximum vibrotactile sensitivity (25-700 Hz). The study optimizes the ratio between the power delivered by the waveform within the region of maximum tactile sensitivity, Q/sub fw/ and the total power delivered by the waveform in one cycle, P/sub T/. The excitatory waveform is composed by a burst of short duty cycle rectangular pulses with an envelope modulating the pulse amplitudes, followed by a recovery time. The study enables, for each pulse frequency, to select an optimum pulse width and an optimum envelope for the rectangular pulses that reduces the total power delivered by the waveform while maintaining the same amount of power within the region 25-700 Hz. Results show that by using an optimum envelope and pulse width, the ratio, Q/sub fw//P/sub T/, can be improved nearly 17% at a pulse frequency of 250 Hz and 1,632% at a pulse frequency of 120 Hz, when compared to a 50% duty cycle and rectangular envelope waveform.

Classification of sleep-waking states using modular neural networks

Applies modular neural network models to classify sleep-waking states in infants. The performances of three connectionist models are compared: (a) a multilayer perceptron (MLP), (b) a mixture of experts (ME) and (c) a fuzzy ganglionar lattice (FGL). We propose a new methodology for enhancing neural classifiers based on input variable selection and confusion error analysis using expert criteria. The ME model was more robust than the MLP and FGL models in the presence of inconsistent or noisy data. Input variable selection and confusion error analysis using expert criteria led to parsimonious models with less parameters and better classification rates.