Citation

Yeong, Chiau Wen (2026) Occlusion aware gait analysis and recovery for gait-Based Parkinson's disease screening. Masters thesis, Multimedia University.

Full text not available from this repository.

Abstract

Parkinson’s disease (PD) is a neurodegenerative disorder characterised by the gradual deterioration of neuronal cells in the brain. As the disease advances over time, patients often experience mobility difficulties. This is due to the basal ganglia, a part of the brain that controls and regulates body movement, which is affected in PD and causes patients to exhibit abnormal gait patterns. PD patients typically exhibit gait abnormalities such as reduced stride length, shuffling steps, and impaired movement coordination. Previous studies have shown the effectiveness of PD classification using vision-based approaches in analysing gait features. However, occlusion remains a challenge when it comes to gait analysis using computer vision. This condition can degrade classification performance as pose estimation may fail to detect keypoints and silhouette representations, such as Gait Energy Image (GEI), which suffer from incomplete body contours. This research aims to develop a low-cost, non-invasive gait analysis framework for PD screening using computer vision techniques under occlusion conditions. To achieve this, this research introduced Dual-OccluGait, a framework to recover occluded data for PD classification. The framework consists of two branches operating on one-dimensional (1D) keypoint and two-dimensional (2D) GEI inputs. In the 1D branch, AlphaPose is utilised to extract skeleton keypoints that capture the essential gait dynamics from the video. When occlusion is detected, a recovery process is carried out using unscented tracking with gated initialisation. In the 2D branch, human silhouettes are processed to obtain GEI, which serve as compact representations of gait patterns over time. Upon detecting silhouette occlusion, a coarse recovery is followed by defect-aware, selective refinement using a Generative Adversarial Network (GAN) to reconstruct occluded body regions. Following occlusion recovery in both branches, dual-branch learning fused both modalities to improve PD classification performance. The dual-modality framework is designed to leverage the strengths of 1D keypoint dynamics and 2D visual-based gait representations, which allow the missing information in one modality to be compensated for by contextual cues from the other. Experimental results demonstrate the effectiveness of the proposed approach. In the 1D keypoints branch, the combination of gated initialisation and unscented tracking for recovering missing keypoints achieved MAE as low as 0.6388. Meanwhile, in the 2D GEI branch, coarse recovery with defect-aware selective refinement achieved a PSNR of up to 31.034 and an MAE as low as 0.002828. DualOccluGait attains an overall classification accuracy of 0.9565 by fusing both branches through dual-branch learning. These highlighted the advantage of integrating gait representations for robust PD gait analysis and classification under occlusion.

Item Type:	Thesis (Masters)
Additional Information:	Call No.: TK7882.P3 Y46 2026
Uncontrolled Keywords:	Pattern recognition systems
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7800-8360 Electronics
Divisions:	Faculty of Information Science and Technology (FIST)
Depositing User:	Ms Nurul Iqtiani Ahmad
Date Deposited:	05 Nov 2026 14:21
Last Modified:	05 Nov 2026 14:21
URII:	http://shdl.mmu.edu.my/id/eprint/17078

Downloads

Edit (login required)