基于多尺度融合的车辆轨迹预测研究OA
Vehicle Trajectory Prediction Based on Multi-Scale Fusion
在当前复杂多变的交通场景下,准确预测各类车辆的未来轨迹对自动驾驶技术的迭代发展至关重要.针对复杂交通环境中存在的车辆交互复杂和轨迹不确定性等问题,文章提出了一种基于多尺度融合的车辆轨迹预测模型(multi-scale fusion vehicle trajectory prediction model,MFVM).该模型主要由时间编码器、多尺度融合模块(multi-scale fusion model,MFM)和MFVM解码器三部分组成.首先,通过时间编码器捕捉长序列的时序依赖关系;其次,采用多尺度卷积提取多粒度车辆轨迹特征,并引入时空双通道注意力策略来完成特征信息的融合;最后,通过全连接层和多层感知机(multi-layer perceptron,MLP)解码器以输出预测轨迹.实验结果表明,相较于现有模型,所提出的MFVM模型在1~2 s的各预测时段内均方根误差降低了0.09~0.19 m,在3~5 s内的误差降低了0.42~0.99 m,充分验证了本模型的轨迹预测性能,尤其在长时域预测中具备显著优势与实用价值.
In the current complex traffic environment,accurately predicting vehicle trajectories is crucial for advancing autonomous driving technology.To address the challenges of complex vehicle interactions and trajectory uncertainty in traffic scenarios,this study proposes a multi-scale fusion vehicle trajectory prediction model(MFVM).The model consists of a tem-poral encoder,a multi-scale fusion module(MFM),and an MFVM decoder.First,the temporal encoder captures long-term temporal dependencies.Next,multi-scale convolutional networks extract multi-granularity trajectory features,which are then fused using a spatiotemporal dual-channel attention mechanism.Finally,the fully connected layer and multi-layer perceptron(MLP)decoder generate predicted trajectories.Experimental results demonstrate that compared to existing models,the MFVM model reduces root mean square error by 0.09~0.19 meters in each prediction interval of 1~2 seconds and by 0.42~0.99 meters within 3~5 seconds,fully validating the trajectory prediction performance of this model,which possesses significant advantag-es and practical value especially in long-term prediction.
朱礼龙;谢进;孙西超
合肥大学 人工智能与大数据学院,安徽 合肥 230601合肥大学 数学与统计学院,安徽 合肥 230601蚌埠学院 数理学院,安徽 蚌埠 233030
交通工程
运动预测车辆轨迹预测特征融合自动驾驶时间编码器
motion predictionvehicle trajectory predictionfeature fusionautonomous drivingtime encoder
《安庆师范大学学报(自然科学版)》 2026 (2)
63-71,9
国家自然科学基金青年科学基金项目(62406095)和安徽省优秀科研创新团队项目(2024AH010017)
评论