Details

Autor(en) / Beteiligte

• Ashwini, K
• Amutha, R

Titel

Compressive sensing based recognition of human upper limb motions with kinect skeletal data

Ist Teil von

• Multimedia tools and applications, 2021-03, Vol.80 (7), p.10839-10857

Ort / Verlag

New York: Springer US

Erscheinungsjahr

2021

Link zum Volltext

Quelle

Alma/SFX Local Collection

Beschreibungen/Notizen

• Limb motion recognition and assessment tasks have become one of the most substantial techniques in clinical environment to evaluate the motor capabilities of neurological patients. However, most of the recognition task are often time consuming and computationally expensive as well. In this paper, we propose a compressive sensing based upper limb motion recognition method using kinect skeletal data. Participants with no form of disabilities were selected for this study. Participants were requested to perform three limb actions and their upper limb motions are captured using kinect sensor. The skeletal data from the kinect sensor chosen for the recognition task is initially converted into RGB image. To alleviate the storage and computational issues, images from skeletal data are further compressed using compressive sensing paradigm. The compressed images are then fed onto a feed forward convolutional neural network for learning and recognizing actions. The recognition task is thus performed directly in the compressed domain avoiding the need for the usage of expensive decompression algorithms. Our proposed method is tested on the acquired data as well as two public benchmark datasets namely MSR Action 3D dataset and KARD dataset. An average recognition accuracy of 96.82%, 94.88% and 97.22% is obtained with our proposed method when tested on acquired dataset, MSR Action 3D dataset and KARD dataset respectively. The proposed method is also compared with several state of the art methods available in literature. Though there is not much increase in the average recognition accuracy of the proposed method on KARD dataset, an increase in average recognition accuracy of around 0.03% to 7.38% is achieved with the proposed method on MSR Action 3D dataset compared to many state of the art methods available in literature.