Efficient Human Activity Recognition System Using Long Short-Term Memory

Human activity recognition (HAR) is a popular and challenging area of research, driven by a diverse set of applications. This paper aims to build a system with the fewest sensors in locations thoughtful enough to distinguish six activities using a Long Short-Term Memory (LSTM) approach to give high performance. We used two wearable Inertial Measurement Units (IMU) sensors with a gyroscope, accelerometer, and magnetometer fixed in thewaist and the right ankle of the subject body. For this purpose, ten random subjects were asked to do six activities (walking, sitting, walking upstairs, standing, walking downstairs, and laying) indoors for data acquisition. Then we analyzed these data and used LSTM to classify the labelled dataset with a different number of hidden units, and the best result was with the number of 150 hidden units. After that, we trained the dataset and registered the test accuracy of 98.44%. Thus, the performance of the proposed method achieved the highest accuracy with low computing costs.