Study On The Model Of Fighting Behavior Detection Based On Random Forest

One of the construction goals of Safe City is meeting the needs of public security management,while taking into account the needs for video surveillance.As the construction of Safe City has entered an intelligent stage,experts and scholars are committed to applying technologies such as big data analysis and machine vision to intelligent video surveillance.Fighting occurs in public places usually threats the safety of life and property.And it may even endanger the social safety and cause a bad social impact.Therefore,fighting behavior detection is one of the key problems in the field of video surveillance.However,the existing fighting behavior detection methods has high detection error rate.Aiming at reducing the detection error rate,this paper proposes a fighting behavior detection model based on the optical flow characteristics.The key of proposing the fighting behavior detection method includes the extraction of features,and the training of machine learning model.(1)The features of fighting behavior in existing papers shows similar changing trend between the fighting behavior and normal behaviors.And it causes the high detection error.Aiming at solving the problem,we considered the optical flow characteristics of the fighting behavior,and then define new features for detection.Including the primary and secondary direction consistency feature,the primary and secondary amplitude direction consistency feature.We have evaluated our features in a portion of CASIA database.Experimental results show that,the features defined in this paper can distinguish between the fighting behavior and normal behaviors effectively in different views,which provides a reliable support for the subsequent detection of fighting behavior.(2)Proposing a detection model based on machine learning algorithms.To improve the accuracy of behavior detection,this paper training a classifier based on the Random Forest algorithm.First,we process the primary and secondary direction consistency feature and the primary and secondary amplitude direction consistency feature in order to generate a feature vector.Then,we train the Random Forest classifier by using Gini-index.With the goal of improving the accuracy of behavior recognition,the parameters of the classifier are optimized.We have evaluated our method in CASIA database and UT-Interaction database,and the 10-fold cross-validation is adopted.Experimental results show that,the features we use have an effective distinguishing ability.Compared with the methods based on optical flow histogram,the proposed method has higher accuracy.Compared with SVM and AdaBoost,the Random Forest classifier we use has higher accuracy,lower false alarm and lower missing alarm.