RAM: Real Time Activity Monitoring with feature extractive training

Activity monitoring systems (AMS) detect actions performed by humans. For an AMS to be effectively deployed in daily life, it should partition sensor data streams in real time and determine what activity corresponds to each partition. In this work, a real time continuous activity monitoring system, named RAM, is proposed. RAM detects simple and composite activities, collecting the data with a single 3D accelerometer to produce a non-invasive solution. Classification module of RAM carries out non-predefined feature extraction and activity detection in a coalesced manner thanks to feature extractive training, whereas the state-of-art classifiers need to be fed with the output of a predefined feature extraction scheme. As being a Support Vector Machines (SVM) inspired solution, RAM fulfils multiclass classification with one-against-pseudo class strategy, without generating hyperplanes. The strength of the proposed model lies in that RAM achieves robustness in terms of inter-activity detection consistency and time efficiency with little overhead. Robustness property offers a potential to reduce the need for re-training an expert system, which faces the problem of growing set of activity classes in the real time activity recognition domain. We compared RAM for a set of hand oriented activities, against 8 different configurations, where SVM and K-Nearest Neighbour (KNN) classifiers are fed with different predefined features. We observed that RAM outperforms these configurations in overall accuracy as well as inter-activity detection consistency. We also presented the results of real tests as a proof-of-concept for transition detection in composite activities.