摘要：

现有出行方式识别模型通常依赖在数据采集设备中插入额外传感器元件或提高设备采样率的方法提升识别准确率，但现实工作环境中提高设备采样率或传感器数量的做法会增加采样设备的能耗，并且设备采样率难以统一也影响了出行方式识别模型的准确率。针对上述问题，提出基于多尺度相位聚合-深层神经决策森林的出行方式识别模型。提取轨迹数据中的多尺度局部和全局特征令牌，采用卷积神经网络层提取令牌间的时空相关性。使用相位检测令牌混合层，动态调整神经网络中令牌与固定权重的关系，捕捉令牌间的相位关系，得到多尺度相位聚合的轨迹表示。利用深层神经决策森林算法，得到出行方式的分类结果。实验结果表明，与基于随机森林的出行方式识别模型相比，所提模型在3种低频重采样数据上的平均识别准确率提升了2.726个百分点，能够更有效地识别出行方式。

关键词: 出行方式识别, 深度学习, GPS数据, 神经决策森林, 卷积神经网络

Abstract:

Existing travel mode identification models usually rely on inserting additional sensors into the data acquisition equipment or increasing the sampling rate of the equipment to improve accuracy. However, in the real production environment, increasing the sampling rate of equipment or the number of sensors will increase the energy consumption of samples. The difficulty of coordinating the sampling rate of equipment also affects the accuracy of the travel mode identification model. To address the aforementioned issues, a travel mode identification model based on Multi-scale Phase-Aggregation deep Neural Decision Forests(MPA-NDF) is proposed. The multi-scale local and global feature tokens are extracted from the trajectory data, and the Convolution Neural Network(CNN) layer is used to extract the spatio-temporal correlation between tokens. The Phase-Aware Token Mixing(PATM) layer is used to dynamically adjust the relationship between tokens and fixed weights in the neural network, thereby capturing the phase relationship between tokens to obtain the trajectory representation of multi-scale phase aggregation. The classification results of travel modes are obtained by using the NDF algorithm. The results show that the average identification accuracy of the proposed model on three kinds of low-frequency resampling data is improved by 2.726 percentage points compared to that of the Random Forest(RF)-based travel mode identification model. Consequently, the proposed model can identify travel modes more effectively.

Key words: travel mode identification, deep learning, GPS data, neural decision forest, Convolutional Neural Network(CNN)