Application Of Deep Learning On Trajectory Prediction

In this thesis,deep learning based object trajectory prediction system was studied,and the algorithm was extended to trajectory extraction and trajectory prediction.Trajectory is extracted by object detection and multi-object tracking.The thesis proposed a structure of multi-object tracking and improved the robustness of object appearance feature.For trajectory prediction,the thesis studied on two scenario: monitoring scenario and driving scenario,and proposed two different encoder-decoder system for both scenario to model the trajectory and predict trajectory of the future,and increase accuracy.The thesis proposed a multi-object tracking structure.To improve the robustness of appearance feature,the process of feature extraction was optimized: to enlarge the training data and enrich the data variations,several datasets was gathered to train the network;to improve the effectness of feature on specific scenario,DGD(Domain Guided Dropout)layer was used in finetune;to make the feature more rubust,metric learning was used to map feature into feature space in which features belong to the same object will be more similar.The thesis improved the feature representation to increase the tracking accuracy.In the domain of trajectory prediction,the thesis studied on monitoring scenario and driving scenario.For monitoring scenario,the thesis introduced an encoder-decoder system with two information paths.First,a LSTM(Long Short-Term Memory)based encoder was used to model the history trajectory,and an attention mechanism was used to lift the weight of those features who are more effective for current prediction.Second,a social encoder was added to the system,the purpose is to help the system learn how can object distribution influence the target moving.For driving scenario,the thesis added a CNN(Convolutional Neural Network)used to extract the image information,the aim is to help the system to percept the surrounding environment of the target.The thesis proposed three information paths to improve the driving scenerio trajectory prediction accuracy.For multi-object tracking and trajectory prediction,a lot of experiments have been carried out.And experiment results prove that the proposed multi-object tracking algorithm and trajectory prediction algorithm can achieve desired effect,the whole system is executable and effective.