Research On Trajectory Prediction Method Based On Deep Neural Network

Time series is a very important data object in life.It exists widely in various fields,including financial data,meteorological observation data,and flying object trajectory data,etc.all exist in the form of time series data.And trajectory prediction can predict the next position information of the target in real time and accurately,which is an important application in time data processing.Trajectory prediction from radar data is an example of this.It is very challenging due to the small sample size,large noise,instability and non-linearity.The rise of deep learning in recent years has brought new solutions to many classic problems,such as target recognition,text translation,and trajectory prediction.Based on the application of deep neural network to radar target trajectory prediction,this thesis studies how to use the combination of long-term and short-term memory network(LSTM)and mixed density network(MDN)to solve several key problems in radar target trajectory prediction.The main research work of this thesis is as follows:(1)A radar target state prediction algorithm based on Bi-LSTM is proposed.The algorithm uses a bidirectional long-short term memory network to learn the time series characteristics of a large amount of training data to achieve accurate prediction of the radar target state.Experiments show that the algorithm has better filtering performance than the traditional non-linear filtering algorithm,and has better adaptability to the initial emission state changes of the radar target.(2)Aiming at the difficulty of gun position extrapolation technology—calculation of projectile launch point at low launch angle,an end-to-end launch point extrapolation method based on LSTM and mixed density network(MDN)was proposed.Experiments under low firing angle conditions show that this method has better effect than the traditional trajectory extrapolation method based on ballistic model and numerical integration.(3)Aiming at the problem of small amount of radar target detection data in some fields,based on Model Agnostic Meta Learning(MAML),a fast(less training steps)and efficient(using a small number of training samples).Comparative experiments prove that the method performs better than ordinary LSTM network prediction algorithms with a small sample size.