Accuracy Improvement of Complex Sensor-based Activity Recognition Using Hybrid CNN
Sakorn Mekruksavanich, Ponnipa Jantawong, Anuchit Jitpattanakul · 2022
Applying smartphones and smartwatches for Human ActivitiesRecognition (HAR) is beneficial for analyzing the participant's dailybehavior and identifying mishaps, such as fall detection, enabling early healthcare monitoring and fall detection. Numerous HAR studies have investigated different recognition models based on sensor data. Nonetheless, this research mainly focused on basic motions such as sitting, standing and walking. In actual settings, human actions are often accomplished in complex ways. This research will examine the recognition of complicated activities using CNN and LSTM networks as deep learning classifiers. Furthermore, this study presented hybrid CNN for enhancing the identification of complicated activities. We evaluated these deep learning classifiers using WISDM-HARB, a publically available benchmark HAR. This dataset used accelerometer and gyroscope data from smartphones and smartwatches to collect sensor data of complicated activities. The experiment demonstrates that all deep learning classifiers recognize basic actions more accurately than complicated ones. Also, it was observed that the hybrid CNN classifier presented had the most fantastic accuracy with 96.25%. According to the research, these classifiers are effective at recognizing basic activities, but their performance decreases as the sophistication of the actions increases. Nonetheless, hybrid CNN models are competent for handling complicated tasks.