Optimizing Activity Recognition Through Dominant Axis Identification in Inertial Sensors

Recent years have witnessed significant growth in sensors-based human locomotion activities recognition due to the availability of low-cost, low-power, and compact sensors and microcontroller units. While significant research has been conducted on human locomotion activity recognition using inertial sensors, most prior studies heavily rely on data from all axes of the sensors. However, the importance of dominant axes in reducing training and inference time has been largely overlooked in these investigations. This letter presents a novel approach, dominant axes-human activity recognition, which aims to identify the dominant axes of inertial sensors to effectively recognize human locomotion activities. The proposed approach effectively reduces both training and inference time while still achieving substantial accuracy. The approach begins with data collection through dedicated smartphone applications and sensory probes. Subsequently, the collected sensory data undergoes preprocessing and annotation for model training. Further, cross-validation is performed during the training phase to determine the dominant axes, leveraging information about the orientation within the dataset. Finally, this work conducts experiments on the collected dataset to assess the approach's efficacy in terms of accuracy and training time.