Human Detection Of Background Model And Action Recognition Of Dense Trajectory In Surveillance Video

With the rapid development of artificial intelligence,intelligent surveillance has become a development trend.As the main object of social activities,it is of great significance to intelligently detect and identify the human in surveillance video.At present,surveillance video mainly relies on manual analysis to extract useful information after the event,which has high labor cost and can not give early warning and alarm in time.In this paper,the research is for human detection and action recognition in surveillance video.The background modeling method is used to achieve human detection,and the dense trajectory method is applied to achieve action recognition.The main research work is as follows:1.The suppression of "ghost" in human detection: To solve the problem of "ghost" caused by object pixels in background model,a universal Two Channel Background Modeling algorithm(TCBM)is proposed to realize background modeling.In channel one,the two-frame difference method is used for pre-detection,and then the bounding box of the human body is estimated by using the idea of mesh topology.Then,the pixels outside the bounding box are added to the background model.In channel two,multi-frame averaging method is used to collect background pixels that can not be collected in channel one.TCBM can eliminate human pixels from the background model and build an effective background model.In this paper,TCBM is used to improve Codebook and Vi Be(Visual Background extractor).The background model of the improved algorithm does not contain human pixels,and no "ghosts" are detected in the test results.Each index(precision,recall,false positive rate,false negative rate)is better than the previous algorithm.2.Decrease the volume of feature descriptor in action recognition: Because the improved Dense Trajectories(iDT)algorithm requires intensive sampling of feature points to extract a large number of feature descriptors,which takes up a large amount of hardware storage resources.To overcome this problem,the number of feature descriptors is reduced by using trajectory deletion,clustering of feature descriptors,and salient feature extraction.Trajectory deletion is used to delete the invalid trajectory in all trajectories.Then feature descriptors are extracted along the remaining trajectory,and the feature descriptors of each action are clustered,and the clustering centers are used to represent the action.Finally,the salient feature descriptors are extracted from the training set.The codebook is trained by the salient feature descriptors,which is used to encode feature descriptors.According to the experiments in KTH and UCFSports,the volume of feature descriptor is decreased by about 80% in both datasets,and the accuracy of action recognition is also increased.Experiments show that the false detection caused by "ghost" can be suppressed with the improved algorithm in the process of human detection.In the process of action recognition,the improved algorithm reduces the number of feature descriptors,which also increases the accuracy of action recognition.Finally,the improved algorithms have a salient effect compared with other advanced algorithms.