Adaptive Orthetics through Intension Detection

Early detection of movement intentions to support active orthoses or exoskeletons.

Artificial intelligence can be used to recognize intentions to move at an early stage - for example via inertial sensors that provide data on acceleration and rotation. By training on typical movement sequences, an AI model learns to deduce the underlying intention during a movement.
Such methods are being researched at the IMT in order to control active orthoses or exoskeletons in a predictive and context-sensitive manner. The aim is to achieve intuitive human-machine interaction in which the technology supports people not only by reacting, but also by anticipating - for example in rehabilitation or in everyday life.

View of the original set-up on the test subject with the IMUs attached to the chest, center of gravity and thighs. The sensors record movement data in real time and are positioned in such a way that they provide precise information on body dynamics.
View of the original set-up on the test subject with the IMUs attached to the chest, center of gravity and thighs. The sensors record movement data in real time and are positioned in such a way that they provide precise information on body dynamics.
Schematic structure of the system: IMUs on the chest, center of gravity and both thighs are connected via SPI to the microcontroller, which forwards the data to a PC for AI-based processing.
Schematic structure of the system: IMUs on the chest, center of gravity and both thighs are connected via SPI to the microcontroller, which forwards the data to a PC for AI-based processing.
Exemplary data set of a forward movement after a start signal to recognize and annotate the movement intention of a test person, displayed over different time windows.
Exemplary data set of a forward movement after a start signal to recognize and annotate the movement intention of a test person, displayed over different time windows.
Architecture overview: Multiple IMU data streams pass through individual chains of CNN, pooling and dense layers. The obtained features are then aggregated by flattening and dense layers to classify human intention.
Architecture overview: Multiple IMU data streams pass through individual chains of CNN, pooling and dense layers. The obtained features are then aggregated by flattening and dense layers to classify human intention.
To the top of the page