Walking Speed Estimation From a Wearable Insole Pressure System Using a Bayesian Regularized Back Propagation Neural Network

Walking speed is an important input to interactive devices in virtual reality, such as omnidirectional treadmills. To improve the immersion in virtual reality, the facility used for speed estimation should be wearable and non-intrusive. An Insole pressure system is an ideal device for such use. By measuring the distribution of plantar3 pressure, walking speed can be estimated from the velocity of the center of pressure (CoP). Although the linear correlation between the velocity of the CoP and the walking speed is good, the accuracy of speed estimation still needs improvement. Besides, the plantar pressure only contains the speed-related information during stance phase. The speed-related information during swing phase cannot be obtained for speed estimation. Thus, using only the plantar pressure information limits the improvement of speed estimation accuracy.

This study proposes a speed estimation method with a Moticon insole pressure system using a Back Propagation (BP) Neural Network. The Moticon insole pressure system consists of an array of 13 capacitive pressure sensors, and a 3-dimensional micro electro mechanical system accelerometer located at the center of the insole. During stance phase, the 3-dimensional accelerations are close to zero, and the plantar pressure dominates the speed estimation process. During swing phase, the plantar pressure is regarded as zero, and the accelerations can be used for speed estimation. To estimate the walking speed, we extracted 25 parameters from the insole system as the inputs to the neural network. The maximum of plantar pressure, the maximum and minimum of plantar pressure derivative in the fore, mid, and rear foot, and stride frequency (total 6 sensors and 19 parameters) were derived from the plantar pressure sensors. The maximum, minimum, and standard deviation of the anterior-posterior and the vertical acceleration (total 6 parameters) were extracted from the accelerometer. The BP neural network comprised of one input layer with 25 nodes, one hidden layer with number of nodes derived from the Kolmogorov theorem, and one output layer with one node of walking speed.

To validate the performance of the proposed method, we tested the data set collected from four subjects walking on a treadmill under seven different speed conditions. 80% of the acquired data was randomly selected for training the network model. The other 20% was used to verify the performance of the prediction model. The results showed good agreement between actual and estimated speeds presenting the linear correlation coefficient r = 0.9935. The proposed method also achieved higher accuracy than these methods including only plantar pressures as inputs.