An Unsupervised Approach for Graph-based Robust Clustering of Human Gait Signatures
Aylin Taştan, Michael Muma, Abdelhak M. Zoubir · 2020
Classification of gait abnormalities plays a key role in medical diagnosis, sports, physiotherapy and rehabilitation. We demonstrate the effectiveness of a new graph construction-based outlier detection method and and the applicability of a new parameter-free clustering approach on radar-based human gait signatures. Micro-Doppler radar-based human gait signatures of ten test subjects for five different gait types consisting of normal, simulated abnormal and assisted walks are clustered using five different clustering algorithms. The proposed algorithm outperforms existing methods both in cluster enumeration and partition and achieves an overall correct clustering rate of 92.8%. The developed method is promising for performing medical diagnosis in a robust unsupervised fashion.