Human Activity Recognition Based on Morphological Dilation followed by Watershed Transformation Method Muhammad Hameed Siddiqi Ubiquitous Computing Lab, Kyung Hee University, Korea siddiqi@oslab.khu.ac.kr Muhammad Fahim Ubiquitous Computing Lab, Kyung Hee University, Korea fahim@oslab.khu.ac.kr Sungyoung Lee Ubiquitous Computing Lab, Kyung Hee University, Korea sylee@oslab.khu.ac.kr Young-Koo Lee Ubiquitous Computing Lab, Kyung Hee University, Korea yklee@khu.ac.kr Abstract—Efficiency and accuracy are the most important terms for human activity recognition. Most of the existing works have the problem of speed. This paper proposed an efficient algorithm to recognize the activities of the human. There are three stages of this paper, segmentation, feature extraction and recognition. In this paper our contribution is in segmentation stage (based on morphological dilation) and in feature extraction stage (using watershed transformation). The proposed algorithm has been tested on six different types of activities (containing 420 frames). The recognition performance of our method has been compared with the existing method using Principle Component Analysis (PCA) to derive activity features. The results of our proposed method are comparable with the existing work. But in-term of efficiency, our algorithm was much faster than the existing work. The average accuracy and efficiency of the proposed algorithm for recognition was 80.83 % and 302.2 ms respectively. Keywords-component; Human activity recognition;computer vision; morphological dilation; watershedding; image segmentation. I. INTRODUCTION Activity recognition means the analysis of motions and behaviors of a human from low-level sensors. Automatic human activity analysis is an increasingly active research area of pattern recognition and in computer vision community and is considered an important problem in the field of the computer vision and pattern recognition. Generally there are two types of activities that a human is performing commonly - Low-level activities, which is also called micro-activities (μ activities) e.g. sitting, standing, walking, running, etc. - High-level activities, which is also called macro-activities e.g. watching TV, playing tennis, ridding bus, etc. The general architecture of high-level activities or macro activities is shown in Fig.1. Some of the existing pattern recognition methods that were used for activity recognition can only utilize labeled activity samples for patterns, although usually large amounts of unlabeled samples exist as they do not need human’s labeling effort [1]. In video the analysis of human activities is a monotonically increasingly important research area from surveillance, security, content-based video retrieval, animation and synthesis. There are several challenges at various levels of processing that are hard to solve, vigorousness against errors at low-level processing, view and rate-constant representation at mid-level processing, and semantic representation of human activities at higher-level processing. The author of [2] presented a comprehensive survey to address the abovementioned problems of representation, recognition, and learning of human from video and related applications. In [3] the authors have presented a novel method to address the activity recognition problem. They represented the activities by feature vectors from Independent Component Analysis (ICA) on video frames, and then based on these features they recognized the human activities by using Hidden Markov Model (HMM) classifier. Many pattern recognition methods have been proposed for the human activity recognition. The important source or origin for activity recognition is the shape information of the frames in the video. The author of [4] proposed an algorithm, which used Independent Component Analysis (ICA) for feature extraction and based on these features, the activity of different types of human were recognized by trained Hidden Markov Models (HMMs). The objective of this paper was to improve the efficiency and accuracy of recognition of the human activities. The proposed algorithm has been tested on six different types of activities, which gave a reliable accuracy and efficiency of activity recognition The rest of the paper is organized as: Section II presents some of related work on human activity recognition. Section III presents methodology of the algorithm. The results are discussed in Section IV. And finally the conclusion and some of the future work are presented in Section V. II. RELATED WORK The two supervised hierarchical models for activity recognition based on motion words was presented by [5]. In their models, there were two aspects, first a ‘visual word’ was obtained from a large-scale descriptors from the whole standing still + Location Time ridding bus High-level activities Low-level activities Fig.1. general architectural model of high-level (macro) activities 2010 International Conference on Electronics and Information Engineering (ICEIE 2010) V2-433     Volume 2