Investigation and Improvement of Posture Detection Algorithm and Its Implementation in Posture Rectification

Chin, Louis Chun Kai (2019) Investigation and Improvement of Posture Detection Algorithm and Its Implementation in Posture Rectification. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text LouisChinChunKai.pdf Restricted to Registered users only Download (3MB)

Abstract

In this era, most of mankind’s activities are carried out on top of a desk, but human never bother to sit with the right posture and this can lead to problems like back pain. In 2013, the number of working days lost due to sickness absence in UK is 131 million days, and 23.66% (Sickness Absence in 2018) of this number is contributed by back pain and neck pain victims. Little attempt has been done to address this problem. In this project, a posture recognition system has been developed which can passively rectify the user’s sitting posture by alerting him/her. A proper and an improper sitting posture might look quite similar to each other in the eye of sensors. Hence, a posture recognition algorithm dedicated for recognizing and differentiating between a proper and an improper sitting posture has been developed. Research and study have been done on work related to the posture recognition system done by other authors. Their methods and solutions are compared and tabulated. In this project, a methodology is proposed to use Kinect for data acquisition, utilizing its skeleton information. MATLAB is used to train and implement the posture recognition algorithm, Support Vector Machine (SVM) and Artificial Neural Network (ANN) are compared and SVM is deemed to be the best classifier in this application. It is found that SVM with linear kernel has the highest accuracy but the input space of training and validating sets are not linearly separable, hence it is speculated that the testing set is biased. Otherwise, RBF and polynomial kernel are deemed to be the best kernel (both have roughly the same average accuracy tested under different situations). Training, validating, and testing data sets were sampled with the methods shown in section 3.3 Feature Extraction from 3 subjects. Different combination of train and test with single/multiple subjects were carried out and the results were studied. Upon detecting an improper sitting posture, the system will trigger a buzzer to alert the user. The focus of this project will be on posture recognition algorithm to ensure successful and accurate posture recognition. Less focus will be put into building the hardware as it is just a prototype and the aim is to demonstrate the idea. At the end of this project, all 3 research objectives were achieved.