Analysis of performance between kinect v1 and kinect v2 for various facial part movements using supervised machine learning techniques
Facial exercises are a series of facial movements such as exaggeration and deflation in upper, middle, and lower face zone for promoting youth and rejuvenating facial muscles. Previously, the effectiveness of facial exercises for rehabilitation and rejuvenation purposes is still controversial due to...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2022
|
| Subjects: | |
| Online Access: | http://umpir.ump.edu.my/id/eprint/37656/1/ir.Analysis%20of%20performance%20between%20kinect%20v1%20and%20kinect%20v2%20for%20various%20facial%20part%20movements%20using%20supervised%20machine%20learning%20techniques.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Facial exercises are a series of facial movements such as exaggeration and deflation in upper, middle, and lower face zone for promoting youth and rejuvenating facial muscles. Previously, the effectiveness of facial exercises for rehabilitation and rejuvenation purposes is still controversial due to lack of quantitative study. However, recent studies have proven the efficacy of the facial exercises through assistive device and guidance from physiotherapist. The process of rehabilitation and rejuvenation through facial exercises is uninteresting and time consuming. Frequently, lack of motivation and patient are the main reasons for treatment failure. Hence, interesting course of facial exercise treatment and encouragement are essential to increase the success rate and effectiveness. The aim of this study is to analyse the performance of Kinect motion sensor version 1 and 2 for development of facial exercise application based on the analysis of face tracking performance. Common 2D cameras are lack of depth information, hence will result in inaccurate facial point detection. In this study, 3D cameras such as Kinect motion sensors version 1 (v1) and version 2 (v2) are used instead of ordinary 2D camera for more accurate facial points detection. Both Kinect sensors are used to apply the Active Appearance Models (AAM) method to detect and extract the facial features. Then, various classic classification methods such as neural network, complex decision tree, cubic support vector machine (SVM), fine Gaussian SVM, fine k-nearest neighbours (kNN), and quadratic discriminant analysis are applied to analyse the detection accuracy of both Kinect motion sensors. After that, the version of Kinect sensor which has better performance is adopted for facial exercises application. The dataset with highest testing accuracy is analysed for constructing rule-based system for the facial exercises application. Eventually, Kinect v2 has outperformed in almost every task by comparing to Kinect v1 except the raising eyebrows exercise. Furthermore, the training and testing accuracies of face data acquired by using Kinect v2 are significantly higher than Kinect v1. The kNN classification is the most suitable method among the applied classification methods as it marks the satisfying training and testing accuracy for both Kinect v2, which are 97.8% and 94.3% respectively. Hence, the correctly predicted dataset using kNN classification method is used for box plot analysis to obtain the threshold parameters, which are mean, lower quartile values, and upper quartile values of the facial part movements. Then, the threshold parameters are used to construct a rule-based system for the real-time facial exercises application. The facial exercises application features with House-Brackmann grading system which gives evaluation of score and grade after completing each set of selected facial exercises. In conclusion, the developed Kinect-based facial exercise application performs successfully and able to give scoring feedback to the user. The application will be able to motivate and encourage users for home facial exercise purpose. |
|---|