Research On Behavior Recognition Under Surveillance Video Based On The Analysis Of Spatial-temporal Relationship

Behavior recognition under surveillance video is one of the research hotspots in the field of computer vision.This thesis studies this problem by discovering the spatial-temporal relationship between frames.First,the behavior recognition technology based on spatial-temporal consistent classification of optical flows for monitoring abnormal behaviors is proposed.Then,for the shadow problem which is,one of the important factors affecting recognition performance,a spatial-temporal Markov random field(MRF)based moving shadow detection model is proposed,which helps to achieve better behavior recognition.(1)Behavior recognition based on the spatial-temporal consistent classification of optical flowsTo distinguish the various behaviors in the video and make timely alarms for the abnormal behaviors,this thesis proposes a behavioral action classification method based on optical flow,which takes multi-frame image as the processing unit and significant region as the subject to computing.The putative areas are first obtained by optical flow estimation and saliency detection,which eliminates the influence of noises.Then the changes of optical flows are decided based on the consistency constraint of spatial-temporal relationship,and taken as a classification feature input to naive Bayes classifier.Further,the existing training samples are compared so that the probability size of the test samples belonging to each category,and the behavior category according to it,are determined.Consequently,the abnormally detection can be fulfilled.Experimental comparisons with several existing behavior recognition methods show that the proposed optical flow classification based behavior recognition method has a good recognition effect for various behaviors.(2)Moving shadow detection based on spatial-temporal Markov random field modelBehavior recognition is easily disturbed by shadows.To this end,this thesis proposes a moving shadow detection method based on spatial-temporal MRF model.The method firstly obtains the initialization result of the shadow seed by simple threshold judgment of color and texture,and then calculates the corresponding shadow area according to the result.The likelihood probability of the shadow feature distance estimate between the backgrounds and the spatial and temporal prior probabilities of the current region compared to the spatial and temporal neighbors by the two comparisons of the shadow features and the area weighted voting scheme,respectively.The MRF integrates three probability statistics into an energy formula,performs iterative optimization,and obtains the optimal result with a certain ending condition,that is,obtains the final shadow detection result.Experiments show that this method can obtain better detection performances and thus can effectively improve the correct rate of behavior recognition results.