Optimizing On-Body Sensor Placements for Deep Learning-Driven Human Activity Recognition

The effective positioning of sensors and the appropriate data merging play a crucial role in advancing systems for recognizing human activities (HAR) using deep neural networks. This study thoroughly evaluates five well-known deep learning structures designed for HAR based on smartphone multi-position sensors. During various activities, the research analyzed data from accelerometers and gyroscopes placed in five different locations on the body. Convolutional neural network (CNN), Long short-term memory (LSTM), Bidirectional LSTM (BiL-STM), Gated recurrent units (GRU), and Bidirectional GRU models were assessed for the classification of activities, considering different sensor combinations across various positions. The study's main findings confirm that achieving accuracy and F1-scores exceeding 99% for the leading BiLSTM, GRU, and BiGRU models is possible by utilizing multi-modal data from well-suited sensor configurations across wrists, arms, pockets, and belts. These recurrent networks consistently outperform CNN and standard LSTM models, which depend on bimodal inputs. Notably, the wrist offers additional motion-sensing capabilities from both the accelerometer and gyroscope, thereby enhancing the fidelity of recognition across multiple positions. The insights derived contribute guidelines for designing robust sensor placements and strategies for fusing data to optimize HAR driven by deep learning, especially on common smartphones. Applications in various domains, such as automated fitness tracking, patient monitoring, and context-aware computing, can benefit from these on-body sensor-based deep models tailored to specific device locations.