Analysis of Activity Recognition and the Influence of Feature Extraction and Selection in an Android Based Device

Sammanfattning: Tracking activities have lately been very popular in smartphones, which requires that the devices are able to classify activities correctly. Since this type of devices has limitations in both power consumption and computational performance, it is important to keep these factors to a minimum. Therefore, the low cost of the accelerometer sensor is a good platform to build a classification on. Although classifications that only use the accelerometer sensor is far from perfected, as far as accuracy is concerned. To obtain a more accurate classification, it would be necessary to dissect the different parts of the classification process, and investigate if any of the parts can be improved. Many studies have been focusing on the different methods of calculating the classification, leading to many different well tested methods. However, very few have investigated the impact features may have on the classification, using the approach ”more is better”. Therefore this work focuses on feature selection combined with modified evaluation methods. Here we show that more features are not necessarily the best solution and that a modified naive evaluation method in most cases are better than a more recognized one. This can affect classifications in the future, especially since fewer features takes less power to compute. This is only the beginning, more studies are needed. We anticipate that our study will be used as a starting point for more in-depth studies in this field.