Research On Target Trajectory Modeling And Migration Identification Methods

Target behavior recognition is one of the most active and extensive research topics in the field of artificial intelligence and cognitive science.Accurate and efficient target behavior recognition can improve the quality of human-computer interaction and promote the development of various sensing applications.Target behavior recognition systems have also been popularized by the increasing number of surveillance cameras.Target behavior recognition is the analysis and recognition of target behavior in video images.The target trajectory can.provide the key information of the target behavior.According to the tracking target,the coordinate sequence of the target can be obtained,and the connection of all coordinates can present the movement trajectory change of.the target.Different movement trajectories can represent different behaviors,so different target behaviors can be identified through the target trajectory.Trajectory modeling and analysis is the most common behavior recognition technique because of the rich behavior information contained in traj ectory.Due to the growing demand for automatic understanding and recognition of target behavior,trajectory modeling analysis has attracted the attention of academia and related industries.In fact,analyzing and understanding the behavior trajectory of the target is also the basic requirement of video indexing,home security and so on.Based on previous researches,this paper analyzes the advantages and disadvantages of various trajectory modeling methods,and determines the Hidden Markov Model(HMM)as the benchmark method for trajectory modeling in this paper.Firstly,the traditional HMM trajectory recognition model was constructed,and baum-welch algorithm was used to train the model continuously until the optimal parameter estimation was obtained.For each trace pattern category,train a corresponding HMM.In the recognition stage,the probability that a particular HMM may produce a sequence of observed eigenvectors is calculated,and the tag with the highest probability is selected as the recognition result.Secondly,in order to improve the performance of cross-perspective target trajectory identification and save the cost of sample collection from different perspectives,this paper proposes to build a migration model of hidden markov model to solve the problem of cross-perspective target trajectory identification.By solving an optimization problem based on the linear regression model to get the source domain feature space and the target domain perspective mapping function between feature space,based on the mapping function optimization migration source domain observation probability parameter,then according to the target domain perspective of a small amount of labeled data to further optimize transition probability source domain and target domain HMM model is obtained.The experimental results show that the HMM trajectory modeling method has a high accuracy and can effectively identify the target trajectory categories,and the HMM simulation trajectory data has a good effect,which is close to the real sample data,indicating that the trajectory identification study of HMM is effective.The HMM migration model performs well in solving the problem of cross-perspective target trajectory recognition.Compared with the traditional HMM trajectory model,the migration optimization of model parameters can significantly improve the recognition accuracy of the target domain model and significantly reduce the amount of annotated data of the target domain perspective.The comparison with the experimental results of various transfer learning methods also proves that the method proposed in this paper is superior in cross-perspective target trajectory recognition,which lays a foundation for the development of trajectory modeling based on transfer learning.