OVERFIT PREVENTION IN HUMAN MOTION DATA BY ARTIFICIAL NEURAL NETWORK
Abstract
Motion analysis has been an active research area for the past decade. Several approaches had been proposed to detect and recognize motion activity for different applications such as motion estimation, modeling, and reconstruction. However, a suitable classifier is required to be embedded with the surveillance system to ensure accurate motion recognition. During these processes, the recognition system compares the captured motion with the motion database in order to recognize the motion activity. However, the classifier can only recognize the motion activities that are closely fit with the database, and overfitting has been an issue in this process. Hence, this paper is aimed at resolving overfitting problem by using Artificial Neural Network (ANN) for motion classification. The motion data was transformed into numerical data with an aid of Kinovea. Data mining software called WEKA was used to perform motion classification. Multi-Layer Perceptron (MLP), which is known as ANN, was modified to recognize different motion activities in the classification process. It was observed that MLP is able to yield classification accuracy of 97.62%. Overfitting issues were also solved by manipulating learning rates in the ANN classifier. A reduced learning rate from 0.3 to 0.1 improved the classification accuracy of jumping motion by up to 12.04%.
Keywords: overfitting, data pre-processing, classification, multi-layer perceptron, recognition, WEKA
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
Papers on acceptance become the copyright of the journal. However, the authors retain publishing rights with respect to Creative Commons Attribution (CC BY 4.0) International License.