Chinese Journal of Computers   Full Text
Title	Discriminative Random Fields for Online Behavior Recognition
Authors	HUANG Tian-Yu1),2) SHI Chong-De3) LI Feng-Xia2) CHENG Cheng2)
Address	1)(School of Software, Beijing Institute of Technology, Beijing 100081) 2)(Beijing Laboratory of Intelligent Information Technology,School of Computer Science,Beijing Institute of Technology, Beijing 100081) 3)(Department of Information Management, Peking University, Beijing 100871)
Year	2009
Issue	No.2(275—281)
Abstract &   Background	Abstract This paper proposes an online behavior recognition based on Discriminative Random Fields. In this model, by incorporating CRF and HCRF, a Frame-HCRF was extended to model behaviors for frames of motion data. The motion intrinsic dynamics are captured by CRF structure as well as extrinsic dynamics between different behaviors by hidden feature functions. This model can accommodate motion data online processing with unknown future frames. The experiments show that the proposed model perform over than HMM, CRF and HCRF for human behavior modeling and recognition. Keywords conditional discriminative models; CRF; Frame-HCRF; behavior models; behavior recognition Background This work is partly supported by the National Natural Science Foundation of China under grant Nos.60773046, the Advanced Research of General Armament Department Foundation of China under Grant Nos. 2220061084, and Beijing Key Discipline Program. Human behavior modeling is the main supporting technique in applications of intelligent Human-Machine interface, Human-Machine training system, intelligent monitoring system, security system etc. It is an important issue belonging to the field of virtual reality, computer graphics and human interaction. By the statistic of recent five-year Siggraph conferences, the study of data-driven virtual human motion has become popular. Most of the research focused on the generative methods, such as hidden-markov models, non-linear dynamics system and their extensions. Caused by the conditional independence assumption and markov distribution assumption, those models can‘t represent the contextual dependences and feature overlapping of time series data such as human motions. Discriminative Models is the most effective approach to solve the dependencies and overlapping problem in machine translation. This paper proposes an online behavior recognition based on Discriminative Random Fields. A Frame-HCRF was extended to model behaviors for frames of motion data. The behavior intrinsic dynamics were captured by CRF structure, and the extrinsic dynamics by hidden feature functions. This model can accommodate motion data online processing with unknown further frames. The experiments prove that the proposed model for human behavior modeling and recognition.