Background
:
Human Activity Recognition is essential for objectively quantifying movement behavior. While the pioneering Acti4 algorithm for thigh-worn accelerometry has been widely adopted, it was developed primarily for researchers and not for collecting large-scale data with high feasibility. To address this, ActiMotus was created as an open-source, Python-based successor to Acti4, building on the same principles. Therefore, this paper presents a validation of the ActiMotus algorithm.
Material and Methods
:
Five publicly available datasets were analyzed, comprising three age groups: children, adults, and older adults. Two different accelerometers were used, SENS motion (12.5 Hz) and Axivity (50 Hz), which were placed on the lateral or anterior thigh in combination with a lower back-worn sensor. Data were collected and evaluated in both laboratory and free-living environments, with video recordings serving as the ground truth. The algorithm’s performance in detecting nine activity types (lying, sitting, standing, shuffling, walking, fast walking, stair climbing, running, and bicycling) was assessed.
Results
:
The algorithm achieved very high overall accuracy in the laboratory environment (>99%) and moderate accuracy in free-living settings, ranging from 78% to 89% across different age groups and sensor placements. In the laboratory, F1-scores exceeded 99% for all observed activities. In free-living conditions, performance varied; however, for sedentary activities (lying and sitting combined), F1-scores consistently surpassed 96% across all datasets.
Conclusions
:
ActiMotus performs well for Human Activity Recognition in both laboratory and free-living environments, comparable to its predecessors, Acti4 and ActiPASS, and offers a solid foundation for assessing movement behavior in large-scale studies.