A Human Action Recognition Method Based On Multi-features Fusion

Being one of the emerging disciplines, computer vision recently developed veryrapidly, and video-based action recognition, the key technology of video analysis andunderstanding, is widely used in robot navigation, video surveillance systems,intelligent transportation, game and entertainment industries. Video, which is moreintuitive in expression and contains much more information compared with the staticimage, is one of the important sources to get information, so it is widely used in themultimedia technology field recently. However, video data is growing at an alarmingrate, and vision-based human action recognition has seen a large increase in activityamong the computer vision community with applications to visual surveillance, videoanalysis and human-computer interaction. Though researchers in related fields havemade certain achievements, the promotion and application of action recognition is stillfaced with many problems. In this thesis we first make a summary on the state of the art,and then do research on the following aspects.Firstly, we analyze the existing works in the field of Action Recognition, andsummarize the disadvantages of the state of art. To overcome the shortcoming of lessexpressive and large amount of data, we propose a new video descriptor based on denseoptical flow and3DHOG. Built in the video cube centered with the track points, thedescriptor get stronger with variation in spatial dimension and temporal dimension, andis more expressive.Secondly, to overcome the shortcoming of sensitive to the camera motion, weintroduce a novel local descriptor called3DHOFG based on HOF descriptor. Thisdescriptor uses the gradient information of optical flow to characterize the actions invideos. Since in most instances of unfixed camera, the camera motion is always slow,which makes the camera motion is excluded by calculate the gradient of motion.Thirdly, we propose a method to combine the motion and static information invideos to make the descriptor more discriminative to characterize human actions. Thestatic features include SIFT feature and color histograms are extract after we get the keyframes. Experimental results on two video datasets show that this method outperformsor reach the same level of the state of art.Finally, the research contents of this paper are summarized and the next research direction is propounded.