Crowd Behaviors Analysis And Abnormal Trajectory Detection Based On Surveillance Data

Crowd behavior analysis,abnormal trajectory detection and trajectory prediction based on video surveillance data,as a cross research direction of pattern recognition,data mining and intelligent transportation,has always been the focus of research.In recent years,crowd behavior analysis,anomaly trajectory detection and trajectory prediction have become a challenging research direction in the field of crowd management,public space design,virtual environment,anomaly detection and intelligent environment.Aiming at the video surveillance data in complex structured scenes,this paper focuses on three aspects of crowd behavior analysis,anomaly trajectory detection and trajectory prediction,and puts forward a complete set of processing framework.The research work of this article includes:Crowd behavior analysis based on the topic model.Trajectory clustering of structured scenes based on the Fast LDA clustering algorithm.In this paper,we proposed a trajectory encoding method named SPE(structure preserving encoding)to encode trajectory firstly,then we use Fast LDA clustering algorithm clustering the encoded trajectories.Based on the results of trajectory clustering,we learning the motion pattern of the crowd trajectory based on the method of statistical learning,we propose local and global motion characteristics statistics as the trajectory's motion pattern.Local motion statistics include the optimal path,and the statistics of global motion characteristics include the densely region of the scene.Abnormal trajectory detection based on local outliers.We improve the Local Outlier Factor(LOF)algorithm,and propose a new outlier detection algorithm K-LOF(KNN-Local Outlier Factor),and use this algorithm to detect the abnormal trajectory.The K-LOF algorithm proposed in this paper firstly use the KNN algorithm to classify the trajectory,and then use the kd tree to optimize the trajectory search,then use the LOF algorithm to calculate the local outlier factor of trajectory,and the kd tree is also used to optimize the search of the trajectory point,according to local anomaly judging whether the trajectory is abnormal.The K-LOF algorithm proposed in this paper can overcome the defects of LOF algorithm that cannot deal with big data and high dimensional data.In terms of crowd trajectory prediction,this paper selects a neural network algorithm,LSTM(Long Short-Term Memory),to predict crowd trajectories,which is the first time to introduce LSTM neural network algorithm to crowd trajectory prediction.The LSTM trajectory prediction framework in this paper consists of the following four functional modules:input layer,hidden layer,output layer and the network training module.The input layer is responsible for handling the trajectory of the original data,the trajectory data can satisfy the network input requirements;hidden layer LSTM structure to construct a single recurrent neural network,responsible for prediction of the trajectory data;the output layer is responsible for providing trajectory prediction results;network training module is responsible for the network training process and to achieve optimization;thus,trajectory prediction can be realized based on the framework proposed in this paper.