Research Of Human Action Recognition Based On HMM

Human action recognition, aiming at how to make machine to know and understand theaction patterns from the visual information, this technology involves multiple disciplines ofartificial intelligence, pattern recognition, computer vision. It has broad application prospectsin many fields. With the rapid development of society, the application requirements mainlyinvolved intelligent video surveillance increase quickly, hence, human action recognitiontechnology has been gradually becoming a hot research topic.Recognition of human action can generally be divided into three steps, including movingobject detection, feature parameters extraction, classification and recognition, this paper focuson classification and recognition stages after introducing the first two parts briefly, namelyhow to create the best model to describe the action and accurately identify action. HiddenMarkov model (HMM) is one of the most commonly used modeling method, the traditionalHMM training method Baum-Welch algorithm is a hill-climbing algorithm that findssolutions with local optimal solution,which will eventually leading the training to get themodel can not accurately describe the action and affecting recognition accuracy finally. Forthis problem, this paper proposes a hybrid training model combining Baum-Welch algorithmand genetic algorithm. The mixed training model can find the global optimum solutionthrough cross using the Baum-Welch algorithm and genetic algorithm,which combingtraditional training algorithm s fast local searching capability and genetic algorithm s globaloptimization capability. Experiments show that HMM trained by the mixed training model hasbetter performance in describing action pattern than the one trained by traditional trainingalgorithm, and thus it obtain a higher level of action recognition accuracy rate.In addition, we also studied the design and implementation of human action recognitiontechnology based on the OMAP3530mini board. We setted up the hardware and softwaredevelopment environment firstly,then we builded a complete embedded Linux operatingsystem, which including the compilation and implementation of1st&2nd boot process, thekernel,the file system,finally we transplanted and achieved recognition algorithm program onthe OMAP3530mini board. These works would like to be a preliminary exploration ofapplying human action recognition technology eventually on embedded portable device.