Cross-domain self-supervised complete geometric representation learning for real-scanned point cloud based pathological gait analysis

Abstract

Accurate lower-limb pose estimation is a prereq-uisite of skeleton based pathological gait analysis. To achievethis goal in free-living environments for long-term monitoring,single depth sensor has been proposed in research. However,the depth map acquired from a single viewpoint encodes onlypartial geometric information of the lower limbs and exhibitslarge variations across different viewpoints. Existing off-the-shelfthree-dimensional (3D) pose tracking algorithms and publicdatasets for depth based human pose estimation are mainlytargeted at activity recognition applications. They are relativelyinsensitive to skeleton estimation accuracy, especially at thefoot segments. Furthermore, acquiring ground truth skeletondata for detailed biomechanics analysis also requires consid-erable efforts. To address these issues, we propose a novelcross-domain self-supervised complete geometric representationlearning framework, with knowledge transfer from the unlabelledsynthetic point clouds of full lower-limb surfaces. The proposedmethod can significantly reduce the number of ground truthskeletons (with only 1%) in the training phase, meanwhileensuring accurate and precise pose estimation and capturingdiscriminative features across different pathological gait patternscompared to other methods.