Human action recognition

Abstract

In the recent years, various computer vision application opportunities such as human action recognition have emerged. It is important that the actions are efficiently classified and identified from video sequences for video analysis. As detecting and understanding such actions would lead to many helpful applications like assisting the sick and security related

surveillance applications. To ensure this, key motion features are extracted using the optical flow algorithm. In order to help identify such actions from features, classification algorithms such as Multilayer perceptron (MLP) and Support Vector Machines (SVM) are needed. This report will focus on Meta-cognitive Radial Basis Function Network and Projection based learning (PBL-McRBFN) algorithm for classification of human actions. McRBFN consists

of two components, cognitive and meta-cognitive. The cognitive component represents knowledge and the meta-cognitive component enables the measured acquisition of knowledge. Meta-cognitive learning emulates human learning by deciding on what-to-learn, when-to-learn and how-to-learn which helps capture knowledge efficiently. The PBL algorithm computes the optimal output weights with the least computation effort in the cognitive component. Using classification algorithms such as LibSVM, Native Bayes and Bayes Net, we are able to present the decision making abilities by comparing the overall and average efficiencies with PBL-McRBFN. To evaluate performance efficiency of the

algorithms mentioned above, Weizmann and KTH datasets from the human action depositories are used as benchmarks. The statistical results have shown that PBL-McRBFN has performed better than the other classifiers as the results reported in the literature review.