Research On Video Object Tracking Algorithm Based On Recurrent Neural Network

The task of Multiple Object Tracking(MOT)can be described as locating multiple objects and maintaining their identities given an input video.The existing MOT method usually identifies the same object and distinguishes others according to their appearance similarity and motion similarity.And the excellent motion similarity can solve the problem of object deformation,object occlusion,and similar appearance.Considering the motion characteristics,the scene information and interaction of objects,this thesis mainly focuses on how to construct a trajectory prediction model using recurrent neural network.And with more accurate and more realistic trajectory predictions,the model obtains more robust motion similarity,which can improve the accuracy of MOT.In order to solve the problem that the traditional linear trajectory prediction model cannot fully describe the complex motion characteristics of pedestrians,this thesis constructs an object trajectory prediction model based on Long Short-Term Memory Networks(LSTM),which can extract robust motion features from the limited historical information of objects.In real scenarios,pedestrian's motion decisions are influenced by other pedestrians and environments.In order to consider these factors,this thesis improves the independent LSTM model to obtain more robust motion similarity,which can identify the same objects and distinguish different objects more accurately.Based on the well trained prediction model with thousands of pedestrian trajectory data,this thesis studies a complete MOT framework.Firstly,based on the detection hypotheses,the appearance model is constructed to generate conservative tracklets.And then the trajectory prediction model is applied to compute motion similarity of tracklets.Finally,considering appearance similarity and motion similarity,the classic network flow is used to complete data association.This thesis shows the evaluations of the proposed method and the comparison with other methods on public MOT benchmark.And then the advantages and disadvantages of the proposed method are analyzed in the end of this thesis.