3D Human Knee Flexion Angle Estimation Using Deep Convolutional Neural Networks

An Anterior Cruciate Ligament (ACL) injury can cause a serious burden, especially for athletes participating in relatively risky sports. This raises a growing incentive for designing injury-prevention programs. For this purpose, the analysis of the drop jump landing test, for example, can provide a useful asset for recognizing those who are more likely to sustain knee injuries. Knee flexion angle plays a key role within these test scenarios. Multiple research efforts have been conducted on engaging existing technologies such as the Microsoft Kinect sensor and Motion Capture (MoCap) to investigate the connection between the lower limb angle ranges during jump tests and the injury risk associated with them. Even though these technologies provide sufficient capabilities to researchers and clinicians, they need certain levels of knowledge to enable them to utilize these facilities. Moreover, these systems demand special requirements and setup procedures which make them limiting. Due to recent advances in the area of Deep Learning, numerous powerful 3D pose estimation algorithms have been developed over the last few years. Having access to relatively reliable and accurate 3D body keypoint information can lead to successful detection and prevention of injury. The idea of combining temporal convolutions in video sequences with deep Convolutional Neural Networks (CNNs) offer a substantial opportunity to tackle the challenging task of accurate 3D human pose estimation. Using the Microsoft Kinect sensor as our ground truth, we analyze the performance of CNN-based 3D human pose estimation in everyday settings. The qualitative and quantitative results are convincing enough to give an incentive to pursue further improvements, especially in the task of lower extremity kinematics estimation. In addition to the performance comparison between Kinect and CNN, we have also verified the high-margin of consistency between two Kinect sensors.