Publication: Inversed unnormalized distance-weighted k-nn: a kinect human motion classification model based on skeletal joints features
Date
2021-12-01
Authors
Lee, Pui Yi
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Human motion (HM) classification remains an active research field due to its wide applications in surveillance, robotics and medical disciplines. HM is commonly recognized from skeletal joint (SJ) coordinates extracted from videos, images sequence, or marker-based motion capture devices. Microsoft Kinect V1 (MKV1) is among the frequently used tool for its low-cost and capable marker-less motion tracks via sensors. Its datasets were captured in image sequence or video output at 30 fps recorded in RGB (red, green, blue) format, depth images, and SJ data (text file) forms. The SJ data includes coordinates for 45 (3D 15 joints) or 60 (3D 20 joints) (features × the total action captured). Previous studies have revealed that MKV1 is a reliable tool having excellent test-retest reliability, and valid for HM patterns assessments based on SJ detection. MKV1 dataset’s big dimensionality requires efficient Feature Selection (FS) analysis to achieve high classification efficiency in a short running time. Nevertheless, no study has applied FS based on Cronbach’s alpha analysis despite wide growth in FS algorithms’ applications. Additionally, previous works which considered the human SJ suffers from low classification accuracy and inconsistent performances across MKV1 datasets. Therefore, the present study aims to assess the reliable SJ indicators captured in MKV1, to determine the optimal subset of SJ for human motion FS, and to develop a human motion classification model of high accuracy and lowest running time. Each dataset’s reliability analysis was established on three indicators: internal consistency, inter-rater reliability, and intra-rater reliability. FS (FSSJFCA) was performed using the exhaustive search strategy and hybrid method. Filter was initially applied to obtained SJ feature combination based on Cronbach’s alpha reliability indicator for subsequent cross-validation mode classification. The Inversed Unnormalized Distance-Weighted k-NN classification algorithm was developed to segregate Kinect SJ data into motions’ classes. The strategy was deployed on three case studies: two datasets from public domains
(UTKinect-Action3D and Florence 3D Actions) and an experimental dataset (USMKinect) involving nine to 16 simple daily HM activities. Findings show that the FSSJFCA improves classification accuracy (reduces total runtime) by 0% (54.63%), 0.1% (25.56%) and 0% (52.69%) in UTKinect-Action3D, Florence 3D Actions, and USMKinect dataset, respectively. The Inversed Unnormalized Distance-Weighted k NN developed achieved classification accuracy above 98% at a low total runtime ((2 - 3 seconds) and (40 seconds) for (31 × 6028) and (31 x 30693) dimension Kinect HM dataset respectively). The newly developed model, a combination of FSSJFCA and Inversed Unnormalized Distance-Weighted k-NN, outperformed 11 existing algorithms applied to UTKinect-Action3D and/or Florence 3D Actions datasets by 0.49% to 17.87% classification accuracy. The newly developed model has high practicability to classify human motion applicable in Artificial Intelligence, Machine Vision, Biomechanics, Image Processing, and Pattern Recognition.