Abstract
While more and more inexpensive devices with embedded sensors are introduced to improve our living, the challenge is to process and analyze large datasets they collect for identification of vital events and activities. Datasets from wearable motion sensors are used to detect and monitor human fall and activities of daily life (ADLs). Existing methods for detection of fall and ADLs from motion datasets employ feature extraction and machine learning, but they have high classification errors. Thus, they produce false alarms for fall and wrong identifications of ADLs. Similar to motion dataset problem, detection of involuntary muscle activities from large EMG datasets (collected from spinal cord injured individual) is a challenging task. Recent studies have developed locations identification algorithms for spasms, motor units, and contractions on individual channel of the EMG datasets. It is important to know how and when repetitive muscle contractions happen in multiple muscles at the same time and is there any any reason for this involuntary co-activity. We demonstrate that the k-means clustering algorithm can semi-automatically extract training examples from motion data. We also propose one- and two- layer classification networks using neural networks and softmax regression. Moreover, we propose a distance measure, called Log-Sum Distance, for evaluating difference between two sequences of positive numbers. We use the proposed Log-Sum Distance measure to develop algorithms for recognition of human activities from motion data. The sequences of m positive numbers for Log-Sum Distance are residual sum of squares errors produced from modeling m motion time-series with multiple linear regression method. To reduce incorrect classification we define a threshold test and use it in our proposed novel algorithm. Log-Sum Distance measure also has been employed to identify the locations for repetitive muscle contractions in one or multiple channels of EMG recordings. We also propose a method to identify the muscle that triggers the first contraction in an identified region. We extract features from EMG data using wavelet filter and decomposing co-variance matrix for eigenvector. Experiments with fall detection, ADLs recognition and monitoring, and repetitive contractions identification methods proposed here show very high accuracy rates with different benchmark datasets. The proposed use of threshold values for classification of activities decreased incorrect classification rates. In summary, this work introduces novel methods and the state-of-the-art development and training of wearable devices for fall and ADLs recognition and monitoring. It also extends the involuntary muscle activities identification across multiple channels.