针对以往单独利用目标检测网络进行机位内保障车辆冲突判定时判据单一以及可靠性较差的问题，提出了一种基于时空金字塔特征的机位内保障车辆行为检测网络STPA-SlowFast。该网络通过金字塔注意力模块，强化了网络在不同尺度通道下的时空特征提取能力并保留了特征的结构信息，使得网络能够更好地捕捉保障车辆运行过程中差异较小的行为变化；针对现有数据中类别样本比例不均的问题，利用Focal Loss对网络损失函数进行改进，平衡网络对于各类别的权重，减缓网络由于训练类别较少产生的精度失真。通过实验验证分析，相较于目前主流的时空行为检测网络，STPA-SlowFast在机位保障车辆运行行为检测精度方面具有一定的提升，平均检测精度为81.46%，在转弯等低样本类别的特殊行为上精度约提升了23.94%，能够显著提升机位保障车辆运行安全。

To address the limitations of using a singular criterion and the reduced reliability observed in the sole application of target detection networks for resolving conflicts among service vehicles on aprons, an apron service vehicle behavior detection network, termed STPA-SlowFast, was proposed. The network incorporated a spatio-temporal pyramid attention module, which was designed to enhance the extraction of spatio-temporal features across multiple scale channels while preserving the structural integrity of these features. This enabled subtle behavioral variations in the operations of service vehicles to be more effectively captured. To mitigate the issue of unbalanced category sample proportions in existing datasets, the network's loss function was refined with Focal Loss, which adjusted the weighting across different categories. This improvement reduced the distortion in detection accuracy caused by sparse training samples in certain categories. Through experimental validation and analysis, STPA-SlowFast was shown to achieve superior performance compared to existing mainstream spatio-temporal behavior detection networks. The proposed network attained an average detection accuracy of 81.46% for operational behaviors of apron service vehicles. Notably, for low-sample categories involving special behaviors, such as turning, the accuracy is increased by approximately 23.94%. These improvements significantly contribute to enhancing the operational safety of apron service vehicles.