Clustering Techniques for Human Posture Recognition: K-Means, FCM and SOM MALEEHA KIRAN 1 , LAI WENG KIN 1 , KYAW KYAW HITKE ALI 1, 2 1 Centre for Multimodal Signal Processing, Malaysian Institute of Microelectronics Systems (MIMOS Berhad) Technology Park Malaysia, 57000 Kuala Lumpur MALAYSIA malehaa.kiran@mimos.my ; lai@mimos.my ; ali.kyaw@gmail.com http://www.mimos.my ; 2 Department of Electrical and Computer Engineering Faculty of Engineering, International Islamic University Malaysia, Jalan Gombak, 53100 Kuala Lumpur, Malaysia http://www.iium.edu.my Abstract: - An automated surveillance system should have the ability to recognize human behaviour and to warn security personnel of any impending suspicious activity. Human posture is one of the key aspects of analyzing human behaviour. We investigated three clustering techniques to recognize human posture. The system is first trained to recognize a pair of posture and this is repeated for three pairs of human posture. Finally the system is trained to recognize five postures together. The clustering techniques used for the purpose of our investigation included K-Means, fuzzy C-Means and Self-Organizing Maps. The results showed that K-Means and Fuzzy C-Means performed well for the three pair of posture data. However these clustering techniques gave low accuracy when we scale up the dataset to five different postures. Self- Organizing Maps produce better recognition accuracy when tested for five postures. Key-Words: - Surveillance systems, posture recognition, Clustering, K-Means, fuzzy C-Means, Self- Organizing Maps 1 Introduction Automated surveillance systems are rapidly becoming a vital tool in providing security. Such systems are increasingly expected not to only scan the images captured, but also to perform some form of real-time analysis on the scene being recorded. As surveillance cameras and video monitors are increasingly being deployed for security purposes, it becomes more and more challenging to adequately monitor and effectively analyze the data resulting from them. Hence the principal goal of automated surveillance system is to automate video monitoring and this is also crucial in combating operator fatigue especially in circumstances where the number of monitors being used for surveillance is huge. At present most automated surveillance systems are being used to analyze events after they have occurred (for example in post-crime scene analysis). However an intelligent security system should be able to preemptively alert security personnel whenever any suspicious behaviour or event is detected by the system. This would mean the system should have the capability to analyze human behavior and to alert the operators accordingly. One key aspect of analyzing human behaviour is correctly recognizing different types of posture. Ideally, we want to recognize postures with only one static camera and in real time. These constraints placed on the system can be rationalized because in majority of applications for video surveillance, only one static camera is used to observe the scene and any analysis of the captured images is done in real time [1]. This would mean the processing speed of the system is also a crucial aspect of such systems. However for our research the first priority is achieving high accuracy of correct posture recognition. Only after a satisfactory level of accuracy has been achieved, the system will be optimized to improve its speed. 2 Problem Formulation Human posture refers to the arrangement of the body and its limbs [2]. There are several agreed types of human postures such as standing, sitting, squatting, lying, kneeling. However for the purpose of our investigation we focus on recognizing five Recent Advances in Signals and Systems ISSN: 1790-5109 63 ISBN: 978-960-474-114-4