Comparison of open-source markerless pose estimation methods in measuring gait kinematics: a preliminary 2D study

In the domain of computer vision, advances in markerless pose estimation through deep learning methods are leading to an increase in accuracy and efficiency in the analysis of human movement, with potential benefits in a wide range of fields. This approach may be particularly useful for gait analysis in clinical settings due to the open-source nature and machine learning techniques rapidly reduce the costs associated of multiple specialized cameras and equipment, and allow for an accessible method while preserving accuracy. Among open-source software, OpenPose is one of the most popular methods which, being based on a general pre-trained model, allows a direct extrapolation of human body coordinates. On the other hand, DeepLabCut is a deep learning technique that allows to train artificial neural networks, starting from a pre-trained model, to recognize the position of manually identified landmarks, developing custom-trained models based on specific dataset. However, although their increasing application in literature, only few studies compared their accuracy in assessing human movement kinematics and, particularly, gait performance. Furthermore, these studies only used DeepLabCut's pre-trained model, without implement a custom-trained model that, conversely, is enable to exploit the main strength of the transfer learning of DeepLabCut. The aim of this study was to evaluate preliminary data on the accuracy of both pre- and custom-trained models of DeepLabCut, to detect the bodyparts location during healthy locomotion in comparison with OpenPose. In addition, it was explored the ability of both DeepLabCut models and OpenPose in measuring spatiotemporal gait parameters in comparison with a force platform system. Fifteen healthy participants walked ten times, at their own pace, along a 5-m indoor walkway, which accommodated for four force platforms (BTS INFINI-T, 1000 Hz). The sagittal plane views of the walking sequence were recorded by an RGB camera (BTS Vixta, 30 Hz, 640 × 480 resolution) synchronized with the force platforms. OpenPose's “Body_25” pre-trained network was employed generating data for 25 keypoints, while DeepLabCut was implemented using “Model Zoo Full_Human” pre-trained network which produced data for 14 keypoints. The custom-trained model was developed by training the artificial network “ResNet-101” with DeepLabCut. For this purpose, 20 frames of each video were manually labelled with 16 keypoints, for a total of 300 annotated images, and a training of 300.000 iterations was performed. For each participant, the output of the three pose estimation approaches, containing the x-y coordinates of the tracked bodyparts, were imported into a Python script for the analysis. Euclidean distances of common keypoints were calculated for each method and the gait parameters as step length, time, speed and cadence were compared with those extracted from the force platforms. Preliminary finding suggested that the trained DeepLabCut's model proved to be more accurate than its pre-trained counterpart and OpenPose, in the reconstruction of bodyparts. By comparing the gait parameters of each approach with the force platforms as a reference system, higher precision of the trained DeepLabCut's network was confirmed. In conclusion, we found trained DeepLabCut's model to be accurate for pose estimation of gait kinematics.