多分支融合注意力机制的车道线检测模型

为了解决目前车道线检测过程中特征融合不充分、检测精确度低和鲁棒性差的问题, 本文提出一种融合多分支结构和 注意力机制的车道线检测模型 (fusion of multi-branch structure and attention mechanism network,FMANet), 图像编码部分采用多 分支结构和注意力机制, 并选择 swish 作为激活函数, 图像解码部分采用跳跃连接结构, 实现跨层特征融合。 本文利用 TuSimple 公开数据集对 FMANet 模型进行评估与验证, 实验结果表明, 本文所提的 FMANet 模型的 mAP 指标接近 97. 25%, 车 道线检测精确度达到 98. 15%, 此外,通过 CULane 数据集验证 FMANet 模型在不同场景下的检测具有更好的鲁棒性。

Lane detection model of multi-branch fusion attention mechanism

In order to solve the problems of inadequate feature fusion, low detection accuracy and poor robustness in current lane detection, this paper proposes a lane detection model called fusion of multi-branch structure and attention mechanism network (FMANet). In the image coding part, fusion of multi-branch structure and attention mechanism is adopted. swish is selected as the activation function, and the image decoding part adopts the jump connection structure to achieve cross-layer feature fusion. In this paper, TuSimple public dataset was used to evaluate and verify the FMANet model. The experimental results show that the mAP index of the FMANet model proposed in this paper is close to 97. 25%, and the lane detection accuracy reaches 98. 15%. In addition, CULane dataset verifies that the FMANet model has better robustness in different scenarios.