Automatic Ankle Angle Detection by Integrated RGB and Depth Camera System.

Area of Science:

• Biomedical Engineering
• Computer Vision
• Machine Learning

Background:

• Depth cameras like Kinect v2 enable 3D human body movement analysis.
• Kinect v2's joint detection accuracy can be limited, particularly for foot and ankle positions.
• Existing methods struggle with incomplete body visibility, especially when only lower limbs are detected.

Purpose of the Study:

• To develop a novel method for accurate human body tracking and 3D movement representation.
• To improve the precision of ankle angle measurements from depth camera data.
• To enable reliable data extraction even when the Kinect v2 device has limited visibility of the full body.

Main Methods:

• Utilized the Kinect v2 device for depth data acquisition.
• Implemented a region-based convolutional neural network (Mask RCNN) for robust body recognition.
• Employed projection lines to enhance the accuracy of ankle angle calculations from raw data.

Main Results:

• Achieved high correlation with gold standard inertial measurement units (IMUs) (Pearson's r = 0.89 ± 0.04).
• Demonstrated improved ankle angle precision, with a mean absolute error (MAE) of 7.5 ± 1.8 degrees at fixed distances.
• Showcased effectiveness in variable distance scenarios (walking test) with MAE = 4.7 degrees.

Conclusions:

• The proposed method significantly enhances the accuracy of human body tracking, especially for lower limb kinematics.
• Mask RCNN and projection lines offer a robust solution for overcoming Kinect v2 limitations in partial visibility scenarios.
• This technique provides a valuable tool for applications requiring precise 3D human movement analysis from depth sensor data.