Investigating classification performance of hybrid deep learning and machine learning architectures on activity recognition

Human activity recognition (HAR) has become a popular field to recognize people's activities from signals obtained using various types of body placed sensors. The increase in the elderly population will increase cognitive and physical decline due to aging and may cause to serious injuries and deaths if immediate assistance is not provided. For this reason, temporal dynamics and important features should automatically be extracted in order to support the daily life of the elderly and to recognize their physical activities correctly and real time. In the study, the classification performances of deep models (CNNLSTM: CNN long short‐term memory network, ConvLSTM: convolutional LSTM, LSTM: long short‐term memory network) and four other machine learning algorithms (SMVs: support vector machines, k‐NN: k‐nearest neighbor classifier, DT: decision tree classifier, ERF: ensemble random forest) which are known to be successful in HAR were investigated. Features used features used in deep models are automatically generated however in machine learning models were generated by hand. The deep models are developed using with a huge set of activities containing 2520 tests. In the tests, each activity of the volunteers was recorded with three axis accelerometer, gyroscope and magnetometer sensors placed in the waist region of body. As a result, ConvLSTM reached the highest accuracy with 99.86% in deep models, while SVMs achieved the highest accuracy in fall detection among machine learning algorithms with 98.47%. When the classification of 36 activities was examined, the highest accuracy was obtained with CNNLSTM and SVMs with 86.94% and 74.58%, respectively. Deep models proposed in this study are considered to be more applicable in real‐world HAR scenarios where sensors' data of indefinite length are obtained.