Transformer Networks for Trajectory Classification

Research related to Trajectory Classification is actively underway, and its application fields are also very diverse. Existing studies related to trajectory classification mainly used RNN-based models such as SimpleRNN, LSTM, GRU, etc. However, these Seq2Seq models cause a bottle neck problem that does not reflect all information when the length of the input sequence increases during the encoding process. Therefore, we propose a Transformer model for more accurate trajectory classification even in situations where the trajectory input sequence is long. As a dataset, we use MNIST stroke sequence dataset, which expresses the stroke of the numbers of the MNIST as a unit vector trajectory. As a result, Transformer achieved comparable performance to LSTM.