Real-time Posture and Activity Recognition by SmartShoe

Automatic recognition of physical activity and postural allocations such as standing or sitting can be used in behavioral modification programs aimed at minimizing static postures. We have developed and validated a footwear-based physical activity monitor (SmartShoe) that can reliably differentiate between most common postures and activities. In previous works, classification was performed using Support Vector Machines (SVM) which are computationally intensive and not well suited for the relatively low resources available on mobile devices such as cell phones. In this paper, we discuss a method for performing automatic posture classification using Artificial Neural Networks operating with fixed point precision arithmetic. The computational time is optimized through application of forward feature selection for determination of the most significant predictors. The method’s performance is analyzed in terms of both throughput and accuracy and we compare it with the SVM classification. The results demonstrate feature selection picked the 61 most significant features out of 153, the proposed methodology reduces the computation time approximately 6000 times (from 1,347ms to 0.22ms) while maintaining comparable classification accuracy (95.2%).