融合Transformer和卷积LSTM的轨迹分类网络

Trajectory classification network fusing Transformer andconvolutional LSTM

为了减少原始轨迹数据的噪声,充分提取轨迹的时空特征,提高基于轨迹数据的交通模式分类精度,提出一种融合堆叠降噪自编码器、Transformer和卷积长短期记忆网络的轨迹分类网络(networks fusing stacked denoising auto-encoder, Transformer and ConvLSTM,SDAETC)。通过堆叠降噪自编码器减少原始轨迹数据中的噪声;利用结合了Transformer的递归图自编码器,提取到更为丰富的时间特征,同时利用特征图自编码器提取空间特征;改进卷积长短期记忆网络,充分提取轨迹中的时空特征,并与提取到的时间特征和空间特征相融合,从而实现交通模式分类。实验结果表明,提出的SDAETC与基线模型相比,在GeoLife和SHL数据集上的准确率分别提升了1.8%和2%。此外,消融实验结果和模型训练时间分析表明,引入堆叠降噪自编码器、Transfomer和ConvLSTM虽然增加了时间消耗,但是对分类精度有积极贡献。

A trajectory classification network SDAETC,which fuses stacked denoising auto-encoder(SDAE),Transformer and convolutional LSTM(ConvLSTM),is proposed to reduce the noise in the original trajectory data,completely extract the spatiotemporal features of the trajectory,and improve performance on trajectory classification.SDAE is used to reduce the noise in the original trajectory data.Abundant temporal features can be acquired by recurrence plots auto-encoders with Transformer,and spatial features can be obtained by feature segment auto-encoder.The temporal features and spatial features are fused with spatiotemporal features,which are extracted by the enhanced ConvLSTM,to perform trajectory classification.Experimental results show that the proposed SDAETC has improved the accuracy by 1.8%and 2%on the GeoLife and SHL datasets,respectively,compared to the baseline model.In addition,the ablation experiments and model training time analysis show that the introduction of SDAE,Transfomer,and ConvLSTM increases time consumption,but has a positive contribution to classification accuracy.