Human Behavior Analysis Based On Video

Human-action analysis based on video has been playing more and more significantrole in the field of human-computer interaction, intelligent monitoring and medicalassistance. It has been paid wide attention to both at home and abroad in recent years.Our paper’s main tasks are as follows:(1) The most popular detection model, Gaussian model, has been adopted formoving targets’ detection. After de-noising and the connected domain’s extraction, thetargets’ external quadrilateral is calculated. When a minority or a single moving target inthe scene, the external quadrilateral can be used for tracking the targets and our resultsshowed it does work.(2) As the human’s clothes are in different colors and its textures are relativelyvague, the main low-level characteristic for identifying the human behaviors in imageare its shape. As many as68kinds of characteristic values of shape is selected andcalculated, which includes: Hu invariant moments, Zernike moments，the similarity ofhorizontal and vertical projection histograms and pairwise geometrical histograms ofsilhouettes, compact, dispersion, aspect ratio, fitting ellipse direction, roundness,symmetry and so on.(3) As the object to be analyzed is video instead of still image, the motion historyimage (MHI) and motion energy images (MEI) are also introduced. MHI and MEI aremainly used for local motion’s detection.(4) In our paper, the random forest, one of the most popular machine leaningmethods has been adopted to classify the human behaviors. Random forests can notonly give high accuracy, but also can handle a large number of characteristics in realtime. An image database has been established for the testing and the results showed thatour classification results are satisfactory.