Research On Trajectory Prediction Algorithm Based On Attention Mechanism

Time series data is an important data type.They can be seen in all aspects of social life,such as stock data,weather data,vehicles trajectory information,pedestrian trajectory information,etc.According to historical information,the forecasting information of the future can help people understand the development law of things and make countermeasures in advance.Therefore,trajectory prediction has important research significance.At present,there are few researches on the trajectory prediction of radar target data.The high-noise,smallsample,and nonlinear radar data has brought huge challenges to traditional radar trajectory prediction methods.In recent years,thanks to the rapid development of network and GPU,deep learning algorithms have been widely used in target recognition,sentiment analysis,image classification,trajectory prediction and other problems.At the same time,the emergence of attention mechanisms has further improved the learning ability of deep learning models.This paper mainly focus on the radar trajectory prediction problem,and proposed a deep neural network based on the attention mechanism,LSTM and encoder-decoder structure.It is applied to the extrapolation of the launch point of the radar target trajectory.At the same time,this paper utilizes Transformer and BERT,which are state-of-the-art method in NLP field,and applies them to the radar target trajectory prediction problem.This paper also constructs a large-scale3 D radar trajectory data set.Multiple experiments on this data set prove that the performance of the method proposed in this paper is better than traditional algorithms and other deep learning algorithms.The main research work of this paper is as follows:(1)A network based on attention mechanism and encoder and decoder is proposed to solve the problem of extrapolation of the launch point of radar target trajectory.The algorithm uses a two-way LSTM network to encode the input trajectory sequence,uses an attention mechanism to perform weighted summation of the encoded information,and uses an LSTM decoder network to decode and predict.The attention mechanism allows the encoder decoder network to encode the deep information of the time series.Experiments show that compared with the traditional nonlinear filtering algorithm,the algorithm has better performance indicators when extrapolating,and it is adaptable to the initial launch state of the radar target.better.(2)Aiming at the trajectory prediction problem of radar target trajectory,we draw lessons from Transformer network and BERT,which are the most advanced solutions in the NLP field,and apply it to the radar target trajectory prediction problem.The Transformer network is a new type of basic network.Compared with the LSTM network,Transformer has a stronger ability to model time series data.As the two-layer representation network of Transformer,BERT is the preferred solution for many large-scale NLP tasks.Multiple comparative experiments show that,under sufficient sample size condition,both Transformer and BERT show better performance than other algorithms.(3)Designed and implemented a radar target extrapolation system simulation platform based on deep learning algorithms and traditional algorithms.Input an observed radar target trajectory into the system,users can select the appropriate algorithm to extrapolate the launch point position of the target trajectory,and generate a corresponding index report.