基于级联网络和图搜索的轨迹模式学习算法

Trajectory Pattern Learning Approach Based on Cascade Competitive Neural Networks and Graph Search Method

提出了一种基于级联竞争神经网络和有向无环图搜索的运动目标轨迹分布模式提取算法。采用级联竞争神经网络提取不同时序处流矢量的分布,根据轨迹点之间的时序转移关系构造有向无环图,通过深度优先搜索来获取轨迹分布的显式表示,并在此基础上给出了一种基于轨迹点对齐的异常轨迹检测方法。构造的级联网络自动隐含了轨迹点的时序关系,可以处理不同长度轨迹模式的学习问题。不同场景的仿真实验表明此方法可以应用于复杂场景下的目标异常行为检测。

A new motion trajectory learning approach was put forward based on cascade competitive neural networks and directed acyclic graph search method. In this approach, the cascade competitive neural networks was trained to discover the distribution of the flow vectors firstly; and then a directed acyclic graph was constructed according to the time relation of the trajectory points; finally, the depth first search method was adopted to obtain the explicit representation of the trajectory pattern. Based on above works, correspondent method was given to detect the abnormal trajectory. The cascade competitive neural networks represent the flow vectors＇ time orders impliedly and can deal with the problem of trajectory pattern learning with different length properly. The simulation results of different scenes demonstrate that the method is effective for anomaly detection in complicated environments.