Research On LSTM-based Multiple Object Tracking And Trajectory Prediction

In this thesis,multi-camera multi-object trajectory tracking and prediction problem are studied.The problems are divided into three parts: multiple object tracking,object re-identification and trajectory prediction.The thesis have proposed new algorithms on these topics and have proven the performance by experiments.In multiple object tracking,the thesis have proposed a new algorithm which is based on LSTM network.The algorithm applies LSTM to position,velocity,interaction and appearance sub-model,and do data association according to similarity scores obtained from the sub-models.In data association,pre-association is performed in order to reduce time consumption,and several thresholds learned by grid search are used to reduce false positives and false negatives so as to improve performance on hard samples.In object re-identification,the thesis have proposed a new algorithm which is based on LSTM network.Single frame LSTM model can learn alignment inside the bound boxes.Multiple frame LSTM model can learn long term dependency among time sequence.Additional attention model is designed to learn sequence level features.The attention model decide the weights of frames with different qualities and therefore improve the re-identification performance.In trajectory prediction,the thesis have proposed a new algorithm which is based on LSTM network.The algorithm follows the sequence-to-sequence structure.It combines trajectory features,background environment constraints and neighbor object constraints to predict more reasonable trajectories.In this thesis,sufficient experiments have been performed to analyze the algorithm and make comparison with state-of-the-art methods.The experiments have proven the performance of proposed algorithms.