Data Acquisition Experiment Protocol

A gait experiment was conducted to study the biomechanical strategies used by able-bodied individuals when walking over a fixed, uneven terrain. The experiment was approved by the University of Manchester Research Ethics Committee (UREC reference 16086). Six able-bodied individuals participated in the gait experiment. They were all male and had no musculoskeletal limitations. The average height and weight of the participant group were 1.7 m (±0.08) and 74.2 kg (±12.2), respectively.

Participants walked at three self-selected speeds, these were slow, normal, and fast. They walked over level-ground and a custom made fixed, uneven terrain shown in Figure 1. Walking over both terrain types made gait pattern comparison possible. Each participant completed 20 walking trials for each speed, over each type of terrain. This resulted in a total of 120 walking trials for each participant.

Level-Ground (left) and Uneven Terrain (right).

Kinematic and kinetic data were recorded for each participant using two 3D AMTI (Watertown, MA, USA) force plates and six Vicon (Oxford, UK) infrared cameras. The force plates were zeroed before conducting the uneven terrain trials to account for the weight of the introduced terrain. The addition of the uneven terrain had no effect on the calibration of the motion capture system. Surface EMG data was also recorded from eight muscles on each leg; namely the tibialis anterior (TA), medial and lateral gastrocnemius (MG and LG), rectus femoris (RF), medial and lateral vastus (VM and VL), biceps femoris (BF), and semitendinosus (SM). The EMG data was recorded using a Delsys Trigno wireless system (Natick, MA, USA). Even though EMG data from eight muscles was recorded, only data from the three lower leg muscles, namely TA, MG and LG, was used for intent prediction. This was done because the upper leg EMG data showed little variation as participants walked on the different terrains.

The three lower leg muscles used were chosen due to their relative size, proximity to the skin and their contribution to ankle-foot motion along both the frontal and sagittal plane. This made it easier and more effective to use surface EMG electrodes. It also minimized the likelihood of signal crosstalk, which tends to occur when measuring activation from muscles that are situated deeper in the body using surface electrodes. However, the shortfall of this approach was that key muscles that contribute more directly to ankle-foot motion along both the sagittal and frontal planes were omitted. These included the soleus, tibialis posterior, peroneus longus, brevis, and tertius muscles. However, as eversion is a combination of foot abduction and dorsiflexion and inversion is a combination of foot adduction and plantarflexion, the three muscles chosen formed a good basis from which to non-invasively explore intent prediction of multi-axial ankle motion.

The EMG data was automatically synchronized with data from the force plates and the motion capture system using the Vicon system. The kinematic and kinetic data were used to identify key phases of the gait cycle, including foot eversion and inversion as participants walked over the uneven terrain. Swing phase was identified based on the activation and deactivation of the two force plates in relation to each other, and also using the motion capture system. Identification of the gait cycle phases made it possible to segment the EMG data for both frontal and sagittal ankle-foot motion and use it to train the prediction approaches that were implemented.

The participants' muscle activation patterns and the magnitude of activation for the respective muscles were calculated from the measured EMG data (De Lisa, 1998). The EMG data was initially bandpass filtered at 20–450 Hz using a Butterworth filter to remove motion artifact and non-physiological signal content. It was then amplitude normalized based on each participant's maximum (isometric) voluntary contraction (MVC) (Yang and Winter, 1984; Halaki and Ginn, 2012). The normalized EMG data was then low pass filtered using a 2nd order recursive Butterworth filter with a cut-off frequency of 20 Hz to ensure motion artifacts were removed (De Luca, 1997).