Improved class binarization model with data oversampling in gait recognition

Gait is a process of a complete cycle of walking that consist of two-step cycles. It can be said that gait has a high degree of biometric which means that every person has its own unique style of walking. Gait recognition using smartphone accelerometer has been widely used in many research and applications due to the cheap assembly, durability and reliability of the Inertial Measurement Unit (IMU) Microelectromechanical System (MEMS) technology. Gait recognition has been used in many areas such as biomechanics, neuro-rehabilitation, sports medicine, security and many others. Latest research achievement in gait recognition approach is the ability to sufficiently recognize a person with small variations and single data enrolment. In the standard gait recognition, there are four main workflows or levels that include data acquisition, pre-processing, features extraction and classification. However, most of the current research is concentrated on the data acquisition and features extractions with a minimal concentration on other workflows, hence the best accuracy is not fully achieved and optimized. In this thesis, we found several problems at the data acquisition stage, pre-processing stage, and classification stage. At the data acquisition stage, gait data is obtained from predefined places such as pocket, pouch, trousers and other parts of the body. However, due to the limitation of the clothes and culture, the mentioned places may not be suitable for smartphone placement. At the pre-processing stage, linear interpolation is widely used by researchers in order to create a fix sampling rate between data points. However, they never examine the best interpolation rate for usage as the rate affects the number of data and this would significantly affect the overall accuracy. At the classification stage, there are two problems that were observed. The first problem is the single classifier mapping applied by the current researchers which are not suitable because the gait recognition involved many classes and possible of overlapped classes boundary is high, hence multiclass classification or binarization of classes should be adopted. However, some researcher does apply one-vs-all (OVA) and one-vs-one (OVO) multiclass methods but the classes are not widely spread and it is not well distributed among class comparison. The second problem in the classification stage is the imbalance class when binarization dataset is performed after the multiclass classification mapping is applied. To overcome the problems mentioned above, we proposed new methods to tackle the problems at the mentioned stages. At the data acquisition stage, we proposed a method that uses hand as the position of the smartphone. At the pre-processing stage, Linear Interpolation Factor Determinator (LIFD) is proposed by using decision tree and cross-validation evaluation in-order to determine the best linear interpolation rate. At the classification stage, we proposed the used of Random Correction Code (RCC) as the main multiclass classifier mapping. RCC is an extension of Error-correcting Output Code (ECOC) that is used for multiclass classification. To tackle the imbalance class problem, a new oversampling method, Self-adjusted Synthetic Minority Over-sampling Technique (SA-SMOTE) is proposed to automatically assign number of samples on the minority class without human intervention. For the experimentation, gait data using hands (HHScD) is collected from 30 subjects with three different poses. Then it is investigated whether it is viable for the gait recognition process. The dataset was compared with the largest gait database from Osaka University (OU-ISIR-2) which the data was captured from smartphone clipped to the waist belt from 408 subjects. Then our proposed methods was applied to the dataset and comparison with the existing method was evaluated. Our experimental results show improvements of the accuracy in comparison with the previous study.

Text FSKTM 2019 7 - ir.pdf Download (1MB)

Actions (login required)

View Item