CCTEG China Coal Research Institute
State Key Laboratory of Intelligent Coal Mining and Strata Control
Engineering Research Center for Technology Equipment of Emergency Refuge in Coal Mine
Beijing Engineering and Research Center of Mine Safe
School of Automation and Electrical Engineering, University of Science and Technology Beijing

目前煤矿井下人员轨迹预测方法中，Transformer与循环神经网络（RNN）、长短期记忆（LSTM）网络相比，在处理数据时不仅计算量小，同时还有效解决了梯度消失导致的长时依赖问题。但当环境中涉及多人同时运动时，Transformer对于场景中所有人员未来轨迹的预测会出现较大偏差。并且目前在井下多人轨迹预测领域尚未出现一种同时采用Transformer并考虑个体之间相互影响的模型。针对上述问题，提出一种基于Social Transformer的井下多人轨迹预测方法。首先对井下每一个人员独立建模，获取人员历史轨迹信息，通过Transformer编码器进行特征提取，接着由全连接层对特征进行表示，然后通过基于图卷积的交互层相互连接，该交互层允许空间上接近的网络彼此共享信息，计算预测对象在受到周围邻居影响时对周围邻居分配的注意力，从而提取其邻居的运动模式，继而更新特征矩阵，最后新的特征矩阵由Transformer解码器进行解码，输出对于未来时刻的人员位置信息预测。实验结果表明，Social Transformer的平均位移误差相较于Transformer降低了45.8%，且与其他主流轨迹预测方法LSTM，S−GAN，Trajectron++和Social−STGCNN相比分别降低了67.1%，35.9%，30.1%和10.9%，有效克服了煤矿井下多人场景中由于人员间互相影响导致预测轨迹失准的问题，提升了预测精度。

Currently, in the prediction methods of underground personnel trajectories in coal mines, Transformer not only has lower computational complexity compared to recurrent neural network（RNN） and long short-term memory (LSTM), but also effectively solves the problem of long-term dependence caused by gradient disappearance when processing data. But when multi personnel are moving simultaneously in the environment, the Transformer's prediction of the future trajectories of all personnel in the scene will have a significant deviation. And currently, there is no model in the field of underground multi personnel trajectory prediction that simultaneously uses Transformer and considers the mutual influence between individuals. In order to solve the above problems, a multi personnel underground trajectory prediction method based on Social Transformer is proposed. Firstly, each individual is independently modeled to obtain their historical trajectory information. Feature extraction is performed using a Transformer encoder, followed by a fully connected layer to better represent the features. Secondly, an interactive layer based on graph convolution is used to connect each other, allowing spatially close networks to share information with each other. This layer calculates the attention that the predicted object allocates to its neighbors when influenced by them, extracts their motion patterns, and updates the feature matrix. Finally, the new feature matrix are decoded by the Transformer decoder to output predictions of future position information. The experimental results show that the average displacement error of Social Transformer is reduced by 45.8% compared to Transformer. Compared with other mainstream trajectory prediction methods such as LSTM, S-GAN, Trajectoron++, and S-STGCNN, the prediction errors are reduced by 67.1%, 35.9%, 30.1%, and 10.9%, respectively. This can effectively overcome the problem of inaccurate prediction trajectories caused by mutual influence among personnel in the underground multi personnel scenario of coal mines and improve prediction precision.

中央高校基本科研业务费专项项目（FRF-TP-24-060A）；天地科技股份有限公司科技创新创业资金专项项目（2023-TD-ZD005-005，2023CG-ZB-10）。

马征，杨大山，张天翔. 基于Social Transformer的井下多人轨迹预测方法[J]. 工矿自动化，2024，50（5）：67-74.

MA Zheng, YANG Dashan, ZHANG Tianxiang. Multi-personnel underground trajectory prediction method based on Social Transformer[J]. Journal of Mine Automation，2024，50（5）：67-74.