ETD

• locomotion mode recognition
• motor primitive
• support vector machines
• wearable robotics

In wearable robotics, identifying the current locomotion is paramount for proper device actuation and user safety. This work proposes a locomotion mode recognition algorithm based on adaptive Dynamical Motor Primitives (aDMP) and Support Vector Machines (SVMs), aimed at initiations of stair descending, stair ascending, and ground-level walking. An IMU sensor was placed on the thigh and used to capture the sagittal roll angle during the swing phase of the initiation movements. The signal was fed to aDMP models to extract features that were given to three SVMs and a custom decision mechanism for generating a prediction. The performance was evaluated offline on ten healthy participants in subject-dependent and subject-independent fashion, resulting in the average accuracies of respectively 98.93% ± 0.70% and 95.43% ± 1.36%, at the heel-strike. Predictions at certain gait events (i.e. velocity zero-crossing and 60% phase of the swing) were also investigated within this work to have a prediction before the next heel-strike, thus, allocating enough time for the system to transition. These predictions led to a decrease in the accuracy, so they were combined in a three-layer classification. Additionally, the algorithm was made possible to imply a safe mode option, in which the wearable device can enter in case of ambiguous locomotion. The preliminary results demonstrate the potential use of this algorithm in transition and steady-state recognition, including for amputee subjects.