Research On Small Cluster Of People Extraction Model Based On Spatio-Temporal Trajectory Representation

With the development and popularization of technologies such as mobile Internet,Internet of Things,camera,social network,and city-aware network,the behavior of users in public places has been effectively recorded,and such time and space information includes the user's travel trajectory and travel intentions.feature.These massive spatial and temporal information can not only portray people's movement patterns,but also become an important part of public safety warning.There are many small groups of people in the society,such as drug abusers,thieves,beggars,unemployed people,etc.These people have different behavioral characteristics than ordinary residents,so there is the possibility of detecting through time and space data.The classification methods based on temporal and spatial trajectory features can be divided into two categories:artificial trajectory feature extraction and end-to-end depth learning.Artificial trajectory feature extraction has three challenges in practical application scenarios:1)In actual research,user behavior data including both space-time dimensions is high-dimensional and sparse;2)In traditional multi-step classifier,trajectory classification is independent of the user classification.The user classification model needs the result of the traj ectory pattern mining as the input parameter.Therefore,the result of the user classification cannot affect the trajectory characteristic parameters.3)In some cases,the number of instances of the positive sample(small cluster of people)is very small and the data set is extremely imbalanced.Therefore,the classification effect of the artificial feature extraction method is not very good.This thesis proposes a scene-based deep learning method to overcome the user identification problem of this small cluster.Firstly,we define the scene of the user's trajectory.We give two kinds of scenes based on the user's history and the place of special meaning.Then,two types of scene modeling based on the Trajectory Representation-based Scenario Model(TRSM)and the Dynamic Graph-based Scenario Model(DGSM)are proposed;a two-layer convolutional neural network was established for user classification.In the experimental part of this thesis,the massive spatiotemporal datasets collected by WiFi sensors are classified.The artificial feature extraction is combined with traditional machine learning methods(including classification model and anomaly detection model).The experimental results show that the model is accurate.Both the precision and the recall rate are superior to other trajectory feature extraction methods.Finally,model-related system prototypes and user interfaces are provided.