Temporal Intelligence: Recognizing User Activities with Stacked LSTM Networks

The advent of sensor data extracted from the smartphone offers opportunities to accurately model the human’s physical activities for context-aware scenarios or maybe for human-centered computing, e.g., human-in-loop, personalized computational tasks, etc. In the paper, we proposed a stacked LSTM-based human activity recognition strategy, which operates on multi-input (sensor datasets, accelerometer, gyroscope, and linear acceleration). The designed activity detection model asserts higher recognition accuracy, 98.56% and significant improvement as compared to other equivalents, over seven human physical activities: Walking, Sitting, Standing, Jogging, Biking, Walking upstairs, and Walking downstairs.