基于交叉注意力的多任务交通场景检测模型

A multi-task traffic scene detection model based on cross-attention

感知是自动驾驶的基础和关键,但大多数单个模型无法同时完成交通目标、可行驶区域和车道线等多项检测任务。提出一种基于交叉注意力的多任务交通场景检测模型,可以同时检测交通目标、可行驶区域和车道线。使用编解码网络提取初始特征,利用混合空洞卷积对初始特征进行强化,并通过交叉注意力模块得到分割和检测特征图。在分割特征图上进行语义分割,在检测特征图上进行目标检测。实验结果表明:在具有挑战性的BDD100K数据集中,所提模型在任务精度和总体计算效率方面优于其他多任务模型。

Perception is the foundation and key issue to autonomous driving,but most single models can’t simultaneously complete multiple detection tasks such as traffic objects,drivable areas,and lane lines.This paper proposes a multi-task traffic scene detection model based on cross-attention,which can detect traffic objects,drivable areas and lane lines simultaneously.Firstly,the encoder-decoder network is used to extract the initial feature maps.Subsequently,the cross-attention module obtains the segmentation and detection feature maps,and hybrid dilated convolution improves the original feature maps.Finally,the semantic segmentation is performed on the segmentation feature maps and object detection is performed on the detection feature maps.The experimental results demonstrate that,on the difficult BDD100K dataset,our model performs better than existing multi-task models in terms of task-wise accuracy and total computational efficiency.