面向城市道路环境时空上下文感知的交通流预测研究

The intelligent transportation system is an effective means to solve traffic congestion, reduce accidents and emissions. Timely accurate traffic flow forecasting is vital for prospective traffic guidance and control, which is one of the classical research topics in intelligent transportation systems. In the previous studies, it is still difficult to estimate the traffic density, and there still exists some possible improvements for traffic flow forecasting for metropolitan road networks and boundaries of metropolitan road networks. This proposal aims to improve the traffic flow forecasting performance on metropolitan road networks, whose objectives include: 1) accurately co-estimating traffic density in surveillance video frames by fusing multi-scale direction-aware attention mechanism and spatiotemporal context features; 2) improving traffic flow forecasting performance by designing a novel metropolitan road network topical-aware attention block to extract topical-aware attention features for co-learning by dilated convolutional neural network and LSTM network; 3) boosting traffic flow forecasting accuracy at boundary points by designing a novel temporal-aware attention block to enlarge the receptive field in temporal context and a loss-switch mechanism to integrate noise-immunity loss function. This project brings new ideas for accurate traffic flow forecasting on metropolitan road networks, so as to improve the quality of road service, and may be widely promoted in different applications.

智能交通是解决交通拥堵、减少事故发生、降低污染排放的有效手段，及时准确的交通流预测是具有前瞻性交通诱导和控制的基础，也是智能交通经久不衰的研究主题之一。在过去研究中，交通流密度的估计仍然存在较大难度，城市道路环境下交通流预测以及城市路网边际点交通流预测的准确性依然存在较大的提升空间。本项目拟以城市道路环境下短期交通流预测为研究目标，主要研究内容包括：1）提出多尺度方向感知注意机制，设计时空上下文特征融合方法，实现低分辨率、高遮挡、大视野的监控视频交通流密度的准确估计；2）构建城市路网拓扑感知注意块，设计有效的拓扑感知注意上下文特征，利用扩张卷积神经网络和LSTM网络进行时空联合学习，提高预测准确性；3）设计时间感知注意机制，扩大时间域感受野，动态切换代价函数，实现噪声免疫的路网边际点交通流的准确预测。本项目的开展将为城市路网交通流预测提供新思路，有助于提升道路服务质量，具有重要应用价值。

智能交通是解决交通拥堵、减少事故发生、降低污染排放的有效手段，及时准确的交通流预测是具有前瞻性交通诱导和控制的基础，也是智能交通经久不衰的研究主题之一。在过去研究中，交通流密度的估计仍然存在较大难度，城市道路环境下交通流预测以及城市路网边际点交通流预测的准确性依然存在较大的提升空间。本项目以城市道路环境下短期交通流预测为研究目标，主要研究内容包括：1）提出多尺度方向感知注意机制，设计时空上下文特征融合方法，实现低分辨率、高遮挡、大视野的监控视频交通流密度的准确估计；2）构建城市路网拓扑感知注意块，设计有效的拓扑感知注意上下文特征，利用扩张卷积神经网络和LSTM网络进行时空联合学习，提高预测准确性；3）设计时间感知注意机制，扩大时间域感受野，动态切换代价函数，实现噪声免疫的路网边际点交通流的准确预测。本项目的开展为城市路网交通流预测提供新思路，有助于提升道路服务质量，具有重要应用价值。

内容获取失败，请点击重试