Overfit Prevention of Human Motion Data By Artificial Neural Network

Low, Kim Yap (2017) Overfit Prevention of Human Motion Data By Artificial Neural Network. Other thesis, INTI INTERNATIONAL UNIVERSITY.

Preview

Text BMEGI 71.pdf Download (640kB) | Preview

Abstract

This report is about overfitting prevention of human motion data. Nowadays, crime is everywhere. To prevent the crimes like burglary, property damage and others rising crimes, a lot of people start to use surveillance system work, the Close Circuit Television (CCTV) will detect and track the human motion by using different method of human motion capturing method such as motion segmentation based on edge detection, and motion recognition. During post-processing stage, the system is compare the human motion data with the others database recognize the motion activity. Sometimes, wrong judgement in post-processing such as misclassification of a running human motion data as walking data or human waving hands data judge as human clapping hands data. All this kind of error call overfitting of human motion data. Objective of this project is to develop a classification method by data Artificial Neural Network (ANN) to prevent the overfitting problem. To achieve my objective, first is to obtain human motion data which transform the video clip data into numerical data, this is because the Waikato Environment for Knowledge Analysis (WEKA) only accept numerical data to perform the classifier. After that, perform data pre-processing which is data imputation to find out the missing value to maximize the efficiency of the output, rearrange all the data in a correct format so that we can use it in WEKA software to perform the classification . In WEKA, it allow us to use Multi-Layer Perceptron classifier to classify our human motion data. The classification accuracy of the input data ranging from 70 to 97%. Which the higher accuracy of human recognition is 97.619% and the motion of the data belong to walking

Actions (login required)

View Item