• No results found

9.1 Classifying handshape

9.1.3 Calibration

The sensors in the CyberGlove are designed to return values in the range 0 to 255. The simplest form of pre-processing for presenting glove data as input to the network is just to scale these inputs to real values in the range desired for the input nodes (which for the networks used in SLARTI was -1 to 1). A more sophisticated approach is to calibrate the data for the range of joint movement of the wearer of the glove. To achieve this it is necessary to perform a series of calibration measurements when the system is started, and record the minimum and maximum values returned by each joint sensor. These ranges are then used during the scaling of future glove data rather than the default range of 0 to 255.

One possible implementation of the calibration process would be to perform it one joint at a time, asking the user to bend and extend the joint to its fullest extent and recording the extreme values produced by this process. This method has two drawbacks. First it involves making two measurements for

each sensor on the glove (36 measurements in all) which makes it a somewhat time-consuming process to have to perform every time the glove is worn. Second the joint ranges measured may be exaggerated relative to the actual range of motion of the joints during natural use of the hand.

Therefore an alternative calibration process was used, wherein the glove's user was asked to make a series of hand gestures intended to capture the normal range of movement of their hand (with emphasis on those features of the hand important for the handshapes used in Auslan). This required only eight measurements and took less than 30 seconds to perform, thereby imposing very little overhead on the use of the CyberGlove. The handshapes used and the features they were intended to measure are described in Table 9.2. The wrist sensors are not of importance with regard to Auslan handshapes, but their calibration was required for use in the orientation and location networks described in Sections 9.2 and 9.3

Table 9.2 Handshapes used for the CyberGlove calibration process

Handshape Features measured

Fist Bending of the finger and thumb joint sensors Spread Extension of the finger, thumb and abduction

sensors

Wish Closure of the index and middle finger abduction sensor

Pinkie and thumb tips touching

Bending of the thumb and pinkie rotation sensors

Wrist flexed up Flexion of the wrist sensors Wrist flexed down Flexion of the wrist sensors Wrist flexed left Flexion of the wrist sensors Wrist flexed right Flexion of the wrist sensors

Scaling the glove inputs relative to these calibrated ranges rather than the gloves default ranges was expected to have a number of benefits. As illustrated in Table 9.3 the actual range of each sensor is much narrower than the default range, and varies considerably between sensors (for example for User A the range of sensor 18 is around 4 times as large as that of sensor 17). Scaling relative to the default ranges would result in considerable variation in the range of values presented to each input node of the network, which may degrade the training process. By scaling relative to the individual calibrated range of each sensor this problem is eliminated.

Table 9.3 also shows that there can be considerable variation in the range of the sensors for different users of the glove. Users A and E (chosen at random) have similar ranges for many of the sensors but are distinctly different for some sensors (such as sensors 1 and 13 which measure rotation of the thumb and pinkie across the palm). This is due to inherent differences in the flexibility and shape of their hands, and also to variations in the fit of the glove. Calibrating for each user should help to reduce these user-specific variations in the data and therefore be of benefit in creating a user independent system.

Due to its flexible construction the CyberGlove does not always fit the user's hand in exactly the same manner from one session to the next. Variations in the data caused by these minor differences in glove positioning can be reduced by performing the calibration process at the start of every session using the glove. Shortening the time required for calibration makes this a feasible proposition.

Table 9.3 Comparison of the calibrated sensor ranges for two different CyberGlove users Sensor # User A min. User A max. User A range User E min. User E max. User E range 1 20 117 97 91 186 95 2 126 186 60 88 179 91 3 94 161 67 76 181 105 4 74 159 85 66 137 71 5 97 185 88 74 181 107 6 121 243 122 130 246 116 7 66 161 95 46 177 131 8 95 195 100 100 196 96 9 173 239 66 166 235 69 10 76 179 103 47 185 138 11 72 184 112 80 182 102 12 124 191 67 115 188 73 13 59 149 90 9 177 168 14 78 219 141 92 219 127 15 139 219 80 102 172 70 16 99 147 48 104 143 39 17 143 184 41 123 189 66 18 69 229 160 44 234 190

25 networks were trained on the raw data, and a further 25 on the calibrated data. The networks used the preferred 16:40:30 topology and were trained for 1,000,000 pattern presentations with a step size of 0.2. The results are summarised in Table 9.4.

Table 9.4 Mean classification accuracy of networks trained using raw and calibrated versions of the handshape data sets

Training set Reg test set Unreg. test set

Raw data 97.9 96.6 87.9

Calibrated data 98.0 96.3 89.9

The results in Table 9.4 show that the performance of the network on the training data was unaffected by the calibration of the glove data. However the generalisation to the test sets was affected. The use of calibrated data slightly reduces the mean performance on the registered test set, but improves the mean on the unregistered signers. Applying a z-test to assess the significance of these differences yields a p-value of 1.88 for the registered signers indicating that there is no significant difference between the calibrated and uncalibrated networks. On the unregistered signers' test set however the p-value is 0.0000 indicating that the calibrated networks' performance on this data is significantly superior to that of the uncalibrated networks. For this reason the calibration process was included in the final SLARTI system.