Academia.eduAcademia.edu

Figure 1 – uploaded by Janez Podobnik

See full PDFdownloadDownload figure
Fig. 1. A subject (left) wearing bands containing inertial measurement units (to] right, actual size 30 mm x 20mm x 5mm) and sneakers containing insoles (botton right).

Figure 1 A subject (left) wearing bands containing inertial measurement units (to] right, actual size 30 mm x 20mm x 5mm) and sneakers containing insoles (botton right).

Related Figures (9)

Fig. 2. Ground reaction force of the left and right foot during gait termination (final step and four preceding steps). Both forces are normalized to the maximum value measured during the trial. Two peaks can be seen for each step, the first representing heel-strike and the second representing toe push-off.
Fig. 2. Ground reaction force of the left and right foot during gait termination (final step and four preceding steps). Both forces are normalized to the maximum value measured during the trial. Two peaks can be seen for each step, the first representing heel-strike and the second representing toe push-off.
Fig. 3. Hip joint angle and ground reaction force of the leading leg during gait initiation, with onset and toe-off marked.
Fig. 3. Hip joint angle and ground reaction force of the leading leg during gait initiation, with onset and toe-off marked.
Mean and median absolute error (AE) for detection of gait onset and toe-off, as well as the percentage of trials where an event was not detected. Results are given separately for within-subject and subject-independent detection. They are further sorted according to type of sensor and trial speed.
Mean and median absolute error (AE) for detection of gait onset and toe-off, as well as the percentage of trials where an event was not detected. Results are given separately for within-subject and subject-independent detection. They are further sorted according to type of sensor and trial speed.
foot. A high GRF suggests that the subject is preparing to push off from the stance foot and make another step, so a low value indi- cates gait termination. If the subject is not leaning forward (left half of tree), the tree checks the standard deviations of the swing leg’s joint angles. A lower standard deviation represents a reduced range of motion during the step, indicating gait termination. Fur- ther branches reevaluate COP and joint angles to resolve uncertain decisions. Fig. 5. It was trained using both normal and fast trials with a buffer length of 0.5s. It initially checks the COP of the stance leg. As the foot is approximately 29cm in length and COP is measured from the toes (0 cm) to the heel, a very low COP means that the subject is leaning forward and the COP has shifted toward the toes, indicat- ing that the subject will continue walking. If the subject is leaning forward (right half of tree), the tree checks the GRF of the stance Fig. 5. A classification tree for detection of gait termination using inertial measurement units and insoles. The tree was trained on both normal and fast trials from all subject with a buffer length of 0.5 s. It represents the optimal subject-independent classification tree when all sensors are available. Features are as defined in Section 2.6. Variable such as knee_swing represent joint angles rather than joint angular velocities (e.g. knee_swing is knee joint angle of the swing leg). Output: no=this is not the last step yes = this is the last step.
foot. A high GRF suggests that the subject is preparing to push off from the stance foot and make another step, so a low value indi- cates gait termination. If the subject is not leaning forward (left half of tree), the tree checks the standard deviations of the swing leg’s joint angles. A lower standard deviation represents a reduced range of motion during the step, indicating gait termination. Fur- ther branches reevaluate COP and joint angles to resolve uncertain decisions. Fig. 5. It was trained using both normal and fast trials with a buffer length of 0.5s. It initially checks the COP of the stance leg. As the foot is approximately 29cm in length and COP is measured from the toes (0 cm) to the heel, a very low COP means that the subject is leaning forward and the COP has shifted toward the toes, indicat- ing that the subject will continue walking. If the subject is leaning forward (right half of tree), the tree checks the GRF of the stance Fig. 5. A classification tree for detection of gait termination using inertial measurement units and insoles. The tree was trained on both normal and fast trials from all subject with a buffer length of 0.5 s. It represents the optimal subject-independent classification tree when all sensors are available. Features are as defined in Section 2.6. Variable such as knee_swing represent joint angles rather than joint angular velocities (e.g. knee_swing is knee joint angle of the swing leg). Output: no=this is not the last step yes = this is the last step.
Fig. 4. A classification tree for detection of gait onset using inertial measurement units. The tree was trained on both normal and fast trials from all subjects and thus represents the optimal subject-independent classification tree for inertial measure- ment units. The joint angles and angular velocities are as defined in Section 2.4. No =onset has not occurred; yes = onset has occurred.
Fig. 4. A classification tree for detection of gait onset using inertial measurement units. The tree was trained on both normal and fast trials from all subjects and thus represents the optimal subject-independent classification tree for inertial measure- ment units. The joint angles and angular velocities are as defined in Section 2.4. No =onset has not occurred; yes = onset has occurred.
Fig. 7. Percentages of correctly analyzed trials when detecting gait termination. Results are shown for different sensors with within-subject and subject- independent cross-validation. Only fast trials are included. Subject-independent validation yields worse results. The median AE when using IMUs, for instance, is over 0.2s for both events. The percentage of undetected events generally remains below 3%, so gait initiation is nonetheless detected. Still, detection algorithms should be trained for each subject separately if possible.
Fig. 7. Percentages of correctly analyzed trials when detecting gait termination. Results are shown for different sensors with within-subject and subject- independent cross-validation. Only fast trials are included. Subject-independent validation yields worse results. The median AE when using IMUs, for instance, is over 0.2s for both events. The percentage of undetected events generally remains below 3%, so gait initiation is nonetheless detected. Still, detection algorithms should be trained for each subject separately if possible.
Fig. 6. Percentages of correctly analyzed trials when detecting gait termination. Results are shown for different sensors with within-subject and subject- independent cross-validation. Only normal-speed trials are included.
Fig. 6. Percentages of correctly analyzed trials when detecting gait termination. Results are shown for different sensors with within-subject and subject- independent cross-validation. Only normal-speed trials are included.
Fig. 8. Percentages of correctly analyzed trials when detecting gait termination. Results are shown for different sensors with within-subject and subject- independent cross-validation. Both normal and fast trials are included.
Fig. 8. Percentages of correctly analyzed trials when detecting gait termination. Results are shown for different sensors with within-subject and subject- independent cross-validation. Both normal and fast trials are included.
Fig. 9. Preliminary example of real-time gait event detection on three trials from a single subject. The algorithms were trained on the 10 subjects in this paper and teste on a new subject. Dashed lines represent hip joint angles while the solid line represents the output of the detection algorithm. The output is as follows: 10=before onse 15 =between onset and toe-off, 20=after toe-off, 25 = gait termination imminent. Please note that while the three trials occurred consecutively, intervals between trial (when the subject returns to a starting position) are omitted to save space.
Fig. 9. Preliminary example of real-time gait event detection on three trials from a single subject. The algorithms were trained on the 10 subjects in this paper and teste on a new subject. Dashed lines represent hip joint angles while the solid line represents the output of the detection algorithm. The output is as follows: 10=before onse 15 =between onset and toe-off, 20=after toe-off, 25 = gait termination imminent. Please note that while the three trials occurred consecutively, intervals between trial (when the subject returns to a starting position) are omitted to save space.
Related topics:
EngineeringMachine LearningGait AnalysisTime FactorsAmbulatory BP monitoringBiomechanical Phenomena
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved