An Efficient CNN-LSTM Approach for Smartphone Sensor-Based Human Activity Recognition System

Human Activity Recognition - HAR is one of the most popular area in the filed of sensor technology and smart learning algorithms. Deep learning algorithms are immensely exploited in HAR systems as it eliminates the need of manual feature engineering. Researchers use normal and hybrid deep learning schemes for training and comparing the models. This paper proposes an efficient CNN-LSTM model for recognising daily human activities using smartphone sensor data. A contemporary CNN-LSTM model is created using time distributed feature extraction layers as it can efficiently handle hierarchical features and can selects the relevant features easily using LSTM memorization scheme. The proposed CNN-LSTM model is compared with two other models - DNN and LSTM in terms of accuracy, precision, recall, F1- score, training loss and computational times. The proposed model managed to outperform other models optimally in all the evaluation metrics. Using holdout training and testing split, the model managed to achieve an average accuracy of 97.609% and 98.69% with relu activation function and 100 training iteration. On validating the different models, the hybrid models takes less computational time and managed to achieve an computational efficiency of (76.23 ± 140.76)% from other models.