基于双流卷积多注意力模型的行人意图识别研究

Research on Pedestrian Intention Recognition Based on Dual-Stream Convolutional Multi-Attention Model

摘要: 识别行人等弱势道路使用者的行为意图是自动驾驶汽车做出有效决策和控制动作保护行人和驾驶者安全的前提。本文设计了一种基于双流结构融合时空特征的行人过街意图识别模型(Dual-stream Convolutional Multi-Attention Model, DCMAM)。基于MobileNet引入空间注意力设计空间流卷积模块；基于膨胀3D卷积网络(Inflated 3D ConvNet, I3D)引入时空和空洞卷积设计时间流卷积模块；基于门控循环单元(Gate Recurrent Unit, GRU)搭建双向GRU网络，捕获时空交互信息；引入注意力机制设计双流融合模块。在数据集JAAD和PIE上的实验证明了模型的有效性，意图识别准确率相较于现有方法提高了7%。集成意图识别模型和硬件平台设计行人意图识别系统，通过实车实验验证了意图识别系统的稳定性和准确性。

Abstract: For autonomous vehicles to effectively make decisions and control actions to ensure the safety of pedestrians and drivers, they must be able to recognize the behavioral intention of vulnerable road users, such as pedestrians. This paper designs a pedestrian crossing intention recognizemodel (Du-al-stream Convolutional Multi-Attention Model) based on the fusion of spatiotemporal features of the dual-stream network structure. Introducing spatial attention based on MobileNet to create the spatial flow convolution module; designing the time Stream convolution module by adding spa-tio-temporal convolution and atrous convolution based on Inflated 3D ConvNet (I3D); a bidirectional GRU network is constructed based on Gate Recurrent Unit (GRU) to capture spatio-temporal inter-action information. The attention mechanism is introduced to design the dual-stream fusion module. Comparative experiments on the datasets JAAD and PIE demonstrate the effectiveness of the pro-posed method, with a 7% improvement in intention recognition accuracy compared to existing methods. Based on a hardware platform integrating an intention recognition network model, a pe-destrian intention recognition system is created. The stability and accuracy of the intention recog-nition system is verified through real vehicle experiments.