Skeletal Data based Activity Recognition System

In recent years, Human activity recognition and research related to it are in high demand owing to its application in various fields such as healthcare systems, assisted living, surveillance etc. Activity recognition system in general aims at identifying activities performed by person in an environment. In this work, skeletal data-based activity recognition system is presented. Microsoft Kinect sensor is a motion capture sensor developed by Microsoft for xbox one. It has become most popular for its effortless operation and low cost. The 3D skeletal joint positions obtained from this kinect sensor are used as raw data for classification purpose. Set of statistical features are computed from these skeletal data. The dimensions of statistical features computed are reduced to eliminate correlated and redundant features among them. Principal component analysis (PCA), a technique used for finding smaller number of uncorrelated data is employed for reducing the feature dimension. The dimensionally reduced features are used as training data for training the classifier. The ability of K-nearest neighbour (KNN) classifier and Support vector machine (SVM) classifier in classifying actions are analysed. The classification method proposed is tested on most popular KARD (Kinect activity recognition dataset) dataset and various classification parameters are computed. Among the two classifiers considered, SVM classifier performed better with an average overall accuracy of 97% which is 3.44% higher than the accuracy produced by KNN classifier.