| Abstrakt: |
Recently, activity detection using sensor data is a widely growing field throughout the world, employing the concepts of artificial intelligence (AI) and deep learning (DL). Fitness tracking, sleep monitoring, fall detection, and smart SOS triggering applications are just a few of the use cases. We can use the built-in mobile sensors in smartphones, such as accelerometers and gyroscopes, to achieve smart human activity detection. To perform real-time activity detection for typical behaviors such as walking, jogging, upstairs movement, downstairs movement, stand-to-sit, sit-to-stand, and fall, we must design an iterative mathematical formulation implemented using a deep learning model. As an input to the proposed model, we will use real-time raw data from the sensors. We used two available datasets: the MobiAct dataset and the SisFall dataset, however owing to a lack of training data, we had to create our unique dataset for activities such as falling and jumping. We constructed a 3D matrix of size (20, 20, 3), scaled it in the range of [1, 255], and sent it to a recurrent convolution neural network-long short term memory (RCNN-LSTM) for feature extraction using a sliding window of 200 rows of time series data. This proposed model has been proven to be most effective in this use case and has a training accuracy of 96.24% and test accuracy of 97.85%. The latency of the model is very low thus making it efficient to be used for real-time human activity detection. This model can be used for real-time activity detection in a variety of use cases and tested to be robust in different places and conditions. [ABSTRACT FROM AUTHOR] |
