The effect of depth data and upper limb impairment on lightweight monocular RGB human pose estimation models.

Abstract

Background and objectives: Markerless vision-based human pose estimation (HPE) is a promising avenue towards scalable data collection in rehabilitation. Deploying this technology will require self-contained systems able to process data efficiently and accurately. The aims of this work are to (1) Determine how depth data affects lightweight monocular red-green-blue (RGB) HPE performance (accuracy and speed), to inform sensor selection and (2) Validate HPE models using data from individuals with physical impairments.

Methods: Two HPE models were investigated: Dite-HRNet and MobileHumanPose (capable of 2D and 3D HPE, respectively). The models were modified to include depth data as an input using three different fusion techniques: an early fusion method, a simple intermediate fusion method (using concatenation), and a complex intermediate fusion method (using specific fusion blocks, additional convolutional layers, and concatenation). All fusion techniques used RGB-D data, in contrast to the original models which only used RGB data. The models were trained, validated and tested using the CMU Panoptic and Human3.6 M data sets as well as a custom data set. The custom data set includes RGB-D and optical motion capture data of 15 uninjured and 12 post-stroke individuals, while they performed movements involving their upper limbs. HPE model performances were monitored through accuracy and computational efficiency. Evaluation metrics include Mean per Joint Position Error (MPJPE), Floating Point Operations (FLOPs) and frame rates (frames per second).

Results: The early fusion architecture consistently delivered the lowest MPJPE in both 2D and 3D HPE cases while achieving similar FLOPs and frame rates to its RGB counterpart. These results were consistent regardless of the data used for training and testing the HPE models. Comparisons between the uninjured and stroke groups did not reveal a significant effect (all p values > 0.36) of motor impairment on the accuracy of any model.

Conclusions: Including depth data using an early fusion architecture improves the accuracy-efficiency trade-off of the HPE model. HPE accuracy is not affected by the presence of physical impairments. These results suggest that using depth data with RGB data is beneficial to HPE, and that models trained with data collected from uninjured individuals can generalize to persons with physical impairments.