Research Of Action Recognition Based On Monocular Vision

Action classification based on monocular vision is an important branch in computer visionfield, which is a process of classifying the human action in the environment of a single camera.Its widespread application prospects and potential commercial value in video intelligentsurveillance, interaction between human and machines, virtual reality, content-based videoretrieval and so on. It is becoming a research focus of many scholars, and it is also one of theforward-looking direction for further research.Because of the influences of many factors, such as the complexity of human actions,diversity of actions, the differences between different persons, changes of view ports andilluminations and so on, lead difficulties to human action classification. Especially undermonocular environment, it needs to consider the impact of the self-shielding and loss of depthinformation. However, the analysis based on the monocular vision has some advantages, such assimple structure, easy calibrated, and avoid the limitation of application. So the article is basedon the analysis of the monocular environment.Moving object extraction is a crucial step in the behavior of classification, the movingobject extraction algorithm are mostly based on the background model object extraction. Thispaper proposed a new algorithm in salient region detection for object tracking, and compared itwith some conventional algorithm based on the salient region detection.This paper spreads out mainly on the foundation of background construction based onGaussian mixture model, optical flow, Principal Components Analysis to analyze the humanaction, which choose a combination of local and global information to describe the behavioralcharacteristics. The target shape is the best way to describe the human action and the optical flowcan well reflect the body-movement information. So this article used the target shape and opticalflow information to describe the local action, and used the relationship between the actionsinformation to describe the global motion. According to the statistics in the experiments, thealgorithm designed for actions classification is feasible when aimed at actions having greatdifferences.