Researches On Behavior Mining For Trajectory Data

With the development of wireless communication,positioning and sensor technology,the acquisition of spatiotemporal trajectory data becomes more and more easy.Spatiotemporal trajectory data is composed of a series of trajectory points including location,time,speed,heading and other information,which contains rich spatiotemporal dynamic information of mobile objects.It not only depicts the behavior mode and activity law of mobile objects,but also reveals the internal mechanism of things evolution.Through data mining,we can find valuable laws and knowledge hidden in track data,which can be widely used in activities recommendation,urban planning,public security,national defense and military and many other fields.So,it has important research value and significance.In this context,this thesis studies the behavior mining technology for trajectory data.The main contents are as follows:(1)Research on quality improvement model of trajectory data.This thesis proposes a trajectory data association algorithm.This method introduces the cosine similarity measure of the trajectory feature vector to better solve the problem of long-term interrupted trajectory association.An adaptive trajectory compression algorithm that retains feature points is proposed.The method uses the compression ratio to automatically control the compression distance threshold,which can maintain the overall contour of the trajectory while compressing the trajectory.(2)Abnormal behavior detection based on deep learning.A trajectory anomaly detection algorithm based on variational self-encoding network is proposed.This method uses GRU as the basic unit of encoding and decoding,the reconstruction probability as the anomaly score,and introduces the proportion of suspected anomalies to adaptively adjust the anomaly judgment threshold.On the simulation test set,the accuracy and recall of the anomaly detection algorithm are higher than 90%,and the real-time detection efficiency is high,which is expected to meet the needs in actual scenarios.(3)General behavior mining based on clustering algorithm.Based on Hausdorff distance,this thesis proposes a new way to measure the trajectory distance,which takes into account the similarity of the spatial position of the trajectory and the similarity of the speed,heading and other characteristics.It can be better applied to the case of incomplete matching of the trajectory.For the density peak clustering CFSFDP,an improved algorithm is proposed,which optimizes the neighborhood search radius with Gaussian kernel function and information entropy,and optimizes the neighborhood search radius with particle swarm optimization algorithm by adjusting the selection of clustering center.This method can accurately mine the regular behavior from a large number of track data.