Pedestrian motion state classification using pressure sensors

Abstract

This paper demonstrates a novel approach for motion classification and analysis using pressure sensors worn by a person. The pressure signal is analysed to search for features corresponding to the motion states, and matched against typical human walking pattern. A prototype system is developed which provides motion classification results in real-time. The motion classification results consists of the number of steps taken by the participant together with the corresponding motion state. The system distinguishes the states associated with a person travelling on a lift, walking on stairs, walking on a flat ground and rest. Data from several participants are collected in a measurement campaign using pressure sensors only, which shows a precision rate of over 90% and a recall rate between 89% and 96%, for the states associated with the movement of participant

On the Challenges and Potential of Using Barometric Sensors to Track Human Activity

“…Some studies have used filtering and signal modeling to overcome the latter problem. For instance, moving average filters over a given time window is widely used for that purpose [71,84,86], followed by other finite impulse response (FIR) filters [87] and infinite impulse response (IIR) filters [76,[87][88][89], such as double exponential smoothing [28,69]. Signal modeling, like the sinusoidal fitting model [90] and sigmoidal nonlinear fitting, is commonly used to increase the contrast of elevation changes [91].…”

“…Barometric pressure is more straightforward than inertial sensors in conveying sensed information due to its fairly direct reading, which greatly simplifies the use of classifiers. The most widely used classifiers are decision trees [28,65,68,82,83,87], support vector machines (SVMs) [75,77,95,96], and threshold-based models [76,91,97,98]. Clustering models, such as hierarchical clustering [73,99] and k-Means clustering [71], Bayesian-based classifiers [72,100,101], LSTM models [28,102], and fuzzy inference models [90] have also been used.…”

“…Similarly to accelerometer data, the signal pattern encoded in a barometer output carries sufficient information to recognize a range of human activities. Ghimire et al [87] observed the change in air pressure when a person walks with hands swinging and used this gait pattern to count steps. The gait pattern is also used to detect the walking class with approximately 95% accuracy [87], similar to the performance of available accelerometer-based recognition methods [63,107], and this accuracy fluctuates for both sensors based on the on-body sensor position.…”

“…Ghimire et al [87] observed the change in air pressure when a person walks with hands swinging and used this gait pattern to count steps. The gait pattern is also used to detect the walking class with approximately 95% accuracy [87], similar to the performance of available accelerometer-based recognition methods [63,107], and this accuracy fluctuates for both sensors based on the on-body sensor position. In this application, the barometric data will also be person-dependent like with an accelerometer, since the gait patterns detected from air pressure changes can vary among the population.…”

“…This can be guaranteed if the sensor data are collected from a controlled environment where no other factors that can be mistaken for VDA occur. As shown by Bollemeyer et al [84], Ghimire et al [87], and Muralidharan et al [68], this can be achieved by limiting the studies to indoors, where only weather, sensor accuracy, and built environment effects impact the barometric readings. Liu et al [104] collected experimental data from outdoors, like mountain climbing, while avoiding activities that cause adverse pressure gradients, like transportation.…”