作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• •    

FMAT:面向飞行区交通的模糊记忆增强轨迹预测算法

  • 发布日期:2026-11-17

FMAT: A Fuzzy Memory-Augmented Trajectory Prediction Algorithm for Airfield Traffic

  • Published:2026-11-17

摘要: 飞行区交通轨迹预测是实现航空器与地面保障车辆运行状态感知、冲突风险预警和智能调度的重要基础。针对现有方法在固定滑动时间窗口下难以充分利用跨时段重复运动模式,监控视频中的遮挡、雨雾、光照变化及检测跟踪误差易造成轨迹抖动和局部缺失,以及仅依赖轨迹序列难以充分表达停机位、服务车道和作业区域等场景约束的问题,提出一种面向飞行区交通的模糊记忆增强轨迹预测算法FMAT。该方法将历史位置序列转换为相对位移序列,利用Transformer编码器提取短期运动动态特征;构建Key-Value形式的轨迹记忆库,以历史轨迹片段编码作为Key、对应未来位移模式作为Value,显式存储可复用的长期运动先验。为降低噪声轨迹对记忆检索的影响,引入高斯模糊隶属度刻画当前轨迹与历史模式之间的软匹配关系,并通过Top-K筛选与加权聚合获得记忆增强表示,抑制低相关记忆项的干扰。同时,采用MAE风格视觉编码器提取与观测时刻对齐的视觉上下文特征,并与轨迹特征和记忆先验共同融入Transformer预测框架,实现多源信息协同建模。实验在Vehicle-CCTV、Airport-CCTV和公开TrajAir数据集上开展,并采用不同随机种子进行多次独立重复训练。结果表明,FMAT在多个数据集的ADE、FDE、MR和LTPA指标上取得最优或接近最优的综合性能;在缺少同步视觉信息的公开数据集上,移除视觉分支后的模型仍保持较好的轨迹建模能力。消融实验进一步验证了轨迹记忆库、高斯模糊隶属度、Top-K加权聚合和视觉上下文建模的有效性,参数敏感性、视觉扰动鲁棒性及推理效率实验表明所提方法具有较好的稳定性、泛化能力和部署可行性。

Abstract: Airfield traffic trajectory prediction is an important basis for operational state perception, conflict risk warning, and intelligent scheduling of aircraft and ground support vehicles. Existing methods based on fixed sliding time windows have difficulty fully exploiting recurrent motion patterns across different time periods. In addition, occlusion, rain, fog, illumination variation, and detection or tracking errors in surveillance videos may cause trajectory jitter and local missing observations, while trajectory sequences alone are insufficient to represent scene constraints such as aircraft stands, service roads, and operation areas. To address these problems, a fuzzy memory-augmented trajectory prediction algorithm for airfield traffic, termed FMAT, is proposed. Historical position sequences are first transformed into relative displacement sequences, and a Transformer encoder is employed to extract short-term motion dynamics. A Key-Value trajectory memory bank is then constructed, where encoded historical trajectory segments are used as Keys and their corresponding future displacement patterns are stored as Values to explicitly preserve reusable long-term motion priors. To reduce the influence of noisy trajectories on memory retrieval, Gaussian fuzzy membership is introduced to characterize the soft matching relationship between the current trajectory and historical motion patterns. Top-K selection and weighted aggregation are further employed to obtain a memory-enhanced representation and suppress interference from weakly related memory entries. Meanwhile, an MAE-style visual encoder extracts visual context features aligned with the observation time, which are jointly integrated with trajectory features and memory priors into the Transformer prediction framework for multi-source collaborative modeling. Experiments are conducted on the Vehicle-CCTV, Airport-CCTV, and public TrajAir datasets with multiple independent runs using different random seeds. Results show that FMAT achieves the best or near-best overall performance on ADE, FDE, MR, and LTPA across multiple datasets. On the public dataset without synchronized visual information, the model without the visual branch still maintains favorable trajectory modeling capability. Ablation studies further verify the effectiveness of the trajectory memory bank, Gaussian fuzzy membership, Top-K weighted aggregation, and visual context modeling. Parameter sensitivity, visual perturbation robustness, and inference efficiency experiments demonstrate the stability, generalization ability, and deployment feasibility of the proposed method.