Single And Interactive Human Behavior Recognition Algorithm Based On Spatio-temporal Interest Point

In recent years, the intelligent video surveillance especially the analysis ofhuman behavior recognition has become a hot research direction, and has a veryimportant significance. In order to prevent and avoid the various emergencies occur inpublic, it is urgently hope that there is a rapid and accurate algorithm of humanbehavior recognition. Focus on human behavior recognition technology, this paperproposes an approach for recognizing human activities based on Spatio-TemporalInterest Point for single behavior recognition and interactive behavior recognition.The algorithm at the same time, At the same time of description the thought ofalgorithm, this paper also introduces some current methods often seen in the field ofhuman action recognition. Most researchers have to analysis the single behaviorrecognition then the double action, when they research the interactive behaviorrecognition. But we think as single behavior the interactive behavior recognition canbe consider as a whole, we don’t need to analysis the two single part individually. Soin this paper we proposes an approach that combine the Spatio-Temporal InterestPoint and the GMM based on EM, thatcan be both applied to single behaviorrecognition and Interactive behavior recognition. This method is divided into threelayers: feature extraction layer, feature representation layer and behaviorrepresentation layer.First the feature extraction layer. We use the Dollar’s method of detectSpatio-Temporal Interest Point that based on Gauss function and Gabor wavelerfunction. This method of Spatio-Temporal Interest Point detection can extract enoughpoints and as this method does not need to track the object, so it is not affected by thetracking effect. In order to reduce the computation complexity of the algorithm, weonly use the three coordinate information of these Spatio-Temporal Interest Points:Spatial information x and y, and temporal information t. Experiments show that thismethod can extract the feature information effectively, and lay a good foundation fornext step.Second, feature representation layer. In this paper we use the concept ofkey-words, that is Spatio-Temporal words. In order to overcome the shortcoming that the traditional K-means method do not considering the probability distribution of dataset that may cause error classification, so Gaussian mixture model(GMM) based onEM estimation was used. Because the probability statistical characteristics of GMMalgorithm, so it can be fully considering the probability distribution ofSpatio-Temporal points. So that Spatio-Temporal words are more accurate, and can bemore effective to describe the probability distribution of Spatio-Temporal points. Itcan also lay a good foundation for next step and improve the accuracy of therecognition of human behavior.Finally, behavior representation layer. As in this paper we consider theinteractive behavior recognition as a whole, so the treatment on Spatio-Temporalpoints can be treated as single behavior. While considering the Spatio-Temporalwords also have distribution characteristic, so we have to GMM clustering performson points of every behavior in the training set for a second time, to get the final GMMmodel that we need.In this paper we uses the Weizmann single behavior database and UT-Interactioninteraction behavior database, and a large number of videos shooting by ourselves, toverify our algorithm. The experimental results on activity datasets show that, thisapproach have a satisfactory identification rate of human activities.