DHARM: recurrent convolutional neural network for real-time activity recognition with smartphone sensor data

Recently, the field of activity recognition using sensor data has been quickly growing worldwide, leveraging the concepts of deep learning (DL) and artificial intelligence (AI) with numerous use cases like fall detection, patient activity tracking, fitness monitoring, gesture detection, etc. The proposed dynamic human activity recognition model (DHARM) is built on a recurrent convolution neural network with long-short-term memory (RCNN-LSTM) trained on a motion sensors based custom collected time series dataset along with two publicly available datasets, MobiAct and SysFall. DHARM is a lightweight model with approximately 43 ms of inference latency, which makes it extensible to edge computing devices like mobile phones, smartwatches, etc. for real-time human activity detection. DHARM has been tested to show a training accuracy of 96.82% and a test accuracy of 98.32%.