Human Actions Recognition And Fight Detection Based On Local Spatial-temporal Features

Human actions recognition based on computer vision is of important researchvalue and application potential and is one of the core technologies in videosurveillance system. Experiments on some famous datasets in human actionsrecognition gives the better results.Local representation of action description is the popular method, this thesisrepresents a new local spatial-temporal feature detection and description. Thismethod first detect SURF interesting points and then delete the points that are onbehalf of background according optical flow magnitude, the remaining points arefinal spatial-temporal interesting points. The local neighbours of interesting pointsare described by SURF descriptor and HOF descriptor.Human actions recognition is achieved with spatial-temporal interesting pointsdetector and descriptor proposed by this thesis and is based on Bow and LDA.Experimental results show that the method of spatial-temporal feature detection anddescription can well recognize actions from Weizmann, KTH, UCF dataset, and isof certain robustness to luminance, occlusion and scale.This thesis presents a new key frames extraction technology and the results onUCF dataset shows that this technology can efficiently reduce computation timeand maintain recognition power. Fight detection is accomplished by key framesextraction and spatial-temporal feature detection and description proposed by thisthesis, the recognition rate is up to95%.