Multimodal deep learning of posture and gait pattern classification

Shamsul Anuar, Akmal Luqman (2024) Multimodal deep learning of posture and gait pattern classification. Project Report. Universiti Teknikal Malaysia Melaka, Melaka, Malaysia. (Submitted)

Text (Full Text) Multimodal deep learning of posture and gait pattern classification.pdf Download (2MB)

Abstract

An individual's gait is a unique characteristic impacted by various variables. Gait patterns, the way a person walks, can be used for verification as they are often unconscious behaviors. Conventional gait pattern studies use machine learning, which requires complex data extraction and achieves limited accuracy with unimodal data. This study employs deep learning to classify gait patterns into walking straight, turning right, and turning left. It also explores the impact of unimodal versus multimodal data for higher accuracy. Advanced algorithms are used to visualize 2- and 3-dimensional gait and posture data in Python, processing the data for deep learning. Gait patterns from 14 participants on different paths along a corridor are extracted. The methodology is divided into three phases: data pre-processing, data processing, and data classification. In data preprocessing, Spyder software is used to visualize each participant's gait and posture frames based on timestamp files from an open-source database. In data processing, the Time Frequency Domain (TFD) method, utilizing Short Time Frequency Transformation (STFT) in Spyder, is chosen to overcome limitations in frequency and time domains. In data classification, results from unimodal and multimodal data using deep learning algorithms, specifically Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN), are compared for different gait situations.

Actions (login required)

View Item