An international team of scientists has developed an artificial intelligence architecture capable of recognising more than 30 types of human activity from sensor data — including everyday movements, physical exercise, and falls — achieving accuracy of up to 98.91% in testing, while running directly on devices such as smartwatches, fitness bands, and medical wearables.
The research involved scientists from Innopolis University and Saint Petersburg Electrotechnical University ‘LETI’ in Russia, working in collaboration with researchers from India and the United States. The findings were published in the peer-reviewed journal Complex & Intelligent Systems. According to the team, the technology addresses a persistent challenge in human activity recognition systems: the need to simultaneously identify rapid movements — such as a fall — and sustained movement patterns that unfold over several seconds, such as distinguishing between walking and running.
To meet that challenge, the researchers designed a two-branch architecture. The first branch employs a temporal convolutional network with causal convolutions to detect local and medium-term patterns. The second combines bidirectional recurrent layers with a channel-attention mechanism, enabling the system to analyse long-range relationships within a movement sequence. According to Dmitry Kaplun, lead mathematical programmer at the AI Research Centre of Innopolis University, the first branch identifies what the user is doing at any given moment, while the second analyses the full context of the movement sequence to anticipate the next activity.
Despite its processing capability, the model was designed to be lightweight, incorporating between 50,000 and 62,000 parameters and requiring between two and 3.5 million computational operations per inference. This allows the model to be deployed on resource-constrained devices without transmitting data to the cloud.
The researchers evaluated the model using three publicly available datasets widely used in human activity recognition research. The first comprises nearly 12,000 records from smartphone accelerometers worn by 30 participants, covering 17 activity types including running, stair climbing, jumping, and various types of falls. On this dataset, the model achieved 97.49% accuracy, surpassing the previous benchmark by 2.49 percentage points. Performance was also more balanced across activity categories, particularly in the identification of less frequent movements such as falls.
On the second dataset — comprising approximately four million records collected via sensors placed on the wrists, chest, and ankles of nine participants — the system reached 98.91% accuracy. The data covered 13 activity classes, including Nordic walking, vacuuming, and cycling.
The third dataset, considered the most challenging due to an uneven distribution of data across activity types, was used to assess the model’s ability to recognise infrequent movements. Even under these conditions, the system achieved 92% accuracy and improved the recognition of rare activities by 10.38 percentage points compared with a previously used model. The researchers also tested the system with sensor data subjected to varying levels of noise. Under light noise conditions, accuracy declined from 91.67% to 88.33%, while moderate noise brought it down further to 83.33%. The authors note that these results indicate a gradual degradation in performance under less controlled conditions.
The technology has potential applications in health monitoring systems, patient recovery tracking, athlete training, and consumer wearables. The researchers also identify possible uses in pedestrian behaviour identification within autonomous transport systems.
Source: Diário Económico
Original article: https://www.diarioeconomico.co.mz/2026/10/01/trends/ai/cientistas-desenvolvem-ia-inovadora-para-relogios/
