With the development of the Beijing Winter Olympic Games,more and more people focus on the ice and snow sports,and begin to pay attention to and pay attention to China’s ice and snow competitions,especially the short track speed skating competitions which are full of ornamental and competitive features.Due to the randomness and variability of the sports behavior and trajectory of short track speed skaters in the competition,it is required that the technology of the intelligent ice and snow field can understand the various sports behaviors of the athletes and accurately predict the sports trajectories of the athletes.However,the related technologies of traditional smart ice and snow fields can only simulate the simple movement patterns between athletes,and are not enough to accurately model the movement trajectories of athletes changing all the time,and it is difficult to predict and optimize the movement trajectories of athletes.To overcome the shortage of existing algorithms,this paper studies the short track speed skater track prediction problem based on deep learning technology and puts forward two prediction models to improve the prediction ability and accuracy.The main contents of this paper are as follows:(1)Given that various algorithmic models based on Long Short-Term Memory Networks(LSTM)have achieved remarkable results in trajectory prediction.However,as the length of the input sequence grows,traditional prediction models are not able to extract the feature vectors accurately,which will eventually affect the performance of the prediction models.To address the above shortcomings,LSTM is applied to the trajectory prediction of short track speed skaters,and the PV-LSTM(Position-Velocity-LSTM)prediction model based on the attention mechanism in the sequence-to-sequence framework(Seq2Seq)is proposed.Since the velocity of short track speed skaters is the key information,the encoded velocity feature information is weighted using the attention mechanism to form a new context vector input to the decoder to accurately predict the future trajectory of the athletes.In order to verify the performance of the model,ablation experiments are conducted on the model,comparison experiments with other network models,etc.The experimental results show that the PV-LSTM model achieves good prediction accuracy for short track speed skater trajectory prediction.(2)It is also considered that short-track speed skaters are constantly interacting spatially in the ice track,and such interactions are often irregular,and the PV-LSTM model does not model such frequent and irregular spatial interactions well.To address this problem,a GTF-LSTM(GAT-Temporal-Features-LSTM)prediction model based on graph attention networks is proposed.Considering athletes as nodes on a graph and edges representing spatial interactions,the greater the weight of edges in aggregating information of node neighborhoods,the stronger the corresponding spatial interactions.Then the temporal correlation in the spatial interaction is extracted by the LSTM layer,which further improves the prediction accuracy.In order to verify the performance of the model,ablation experiments are conducted on the model,and comparison experiments with other network models are performed,etc.The experimental results show that the newly proposed GTF-LSTM prediction model has excellent prediction performance under each index. |