COMPARATIVE STUDY OF YOLOV7 AND OPENPOSE FOR GAIT ANALYSIS WITH MARKER METHOD AS BASELINE

• Main Author: Thirafi Hugo Mafaza, Adonis
• Format: Final Project
• Language: Indonesia
• Subjects: Teknologi
• Online Access: https://digilib.itb.ac.id/gdl/view/79476
• Institution: Institut Teknologi Bandung

Research at the ITB Biomechanics lab has extensively utilized OpenPose for markerless motion capture in gait analysis at speeds of 4, 6, and 8 km/hour. A study by Farian reported substantial errors in OpenPose's tracking, with hip joint position errors up to 80 mm horizontally and 45 mm vertically, affecting angular measurements such as thigh, leg, and knee angles by up to 40 degrees. These errors increased with walking speed. Alvin Wen's subsequent fine-tuning of OpenPose with lab data yielded a persistent 17-pixel error. Given these issues, this research explores YOLOv7 as a possible better alternative. This thesis compares YOLOv7 and OpenPose by focusing on knee angle accuracy against the HSV motion capture method, incorporating 2D-DLT camera calibration. The comparative analysis, based on RMSE and Pearson correlation, shows YOLOv7 and OpenPose to have a negligible difference in correlation 0.99 and 0.98 but a significant difference in RMSE 2.24 and 3.30 degrees. An independent two-sample Student’s T-Test confirmed the statistical significance of this difference. In conclusion, YOLOv7 slightly outperforms OpenPose in gait analysis, evidenced by Pearson correlation, RMSE, and t-test results. It is recommended that YOLOv7 be further refined using ITB lab data, as its architecture is conducive to efficient training, even with limited datasets.