The Effect of Window Length on the Classification of Dynamic Activities through a Single Accelerometer

This paper investigates how different window sizes for feature extraction and classification affect the accuracy of daily living locomotors activity recognition through accelerometers. A comprehensive data set was collected from 9 healthy subjects performing walk, stair descending and stair ascending while carrying an accelerometer on the waist. Nearest neighbor based classification has been used because of its simplicity and flexibility. The findings show that, by increasing window length, the system accuracy increases, but it produces delays in real time detection/alert of the activity. From the experiments it is concluded that a 2 seconds (2 s) time window may represent a trade-off for the detection of these mentioned activities in a real-time scenario, as it produces 91.7 percent of accuracy.