基于M形深度架构的非结构化道路可行驶区域推荐模型

针对边界模糊、路况多变的非结构化道路,为满足智能汽车在正常、应急等复杂行驶工况下对可行驶区域的视觉检测需求,提出一种在M形深度架构下融合多尺度交互策略和双重注意力机制的可行驶区域推荐模型,能够在复杂驾驶场景中精细分割出非结构化道路的强推荐、弱推荐、不推荐行驶区域。首先,在编码器-解码器的骨架基础上,构建倒金字塔式的多尺度分层输入和分层输出结构,以有效融合非结构化道路的浅层形态学特征与深层语义信息,并平衡模型在不同尺度上的预测偏倚,提升复杂驾驶场景下对多尺度与变尺度目标的分割精度;其次,构建集成通道注意力和空间注意力的跳跃连接结构,使模型在实现编码特征与解码特征高效传递的同时,聚焦于学习与道路可行驶性相关的重要特征,进一步强化模型对非结构化道路的检测性能。通过多种途径构建包含城郊、乡村、园区等真实场景的非结构化道路驾驶数据集。试验结果表明：得益于M形深度架构对多尺度交互策略和双注意力机制的融合,提出的模型在多种真实驾驶场景下均能较好地实现强推荐行驶区域、弱推荐行驶区域、不推荐行驶区域和背景区域的精细分割,平均交并比达到92.46%,平均检测速度达到22.7帧·s^-1;与现有其他主流模型相比,提出的模型兼顾了分割精度和时间效率,在非结构化道路可行驶区域检测任务上有明显优势。

Drivable Region Recommendation Model for Unstructured Road Based on M-shaped Deep Architecture

For unstructured roads with ambiguous boundaries and changeable traffic, intelligent vehicles have vision-based drivable region detection requirements under normal and emergency driving conditions. In this study, a novel drivable region recommendation model based on an M-shaped deep architecture equipped with multi-scale interaction strategy and dual attention mechanism is proposed to address this problem. The proposed model can accurately extract the drivable region from a complex driving scene and finely divide it into highly recommended, low-recommended, and non-recommended drivable regions. Based on an encoder-decoder backbone, multi-scale hierarchical input and output with an inverted pyramid structure was first proposed to effectively fuse shallow morphological features with deep semantic information of an unstructured road, balance the prediction bias at different scales, and improve the segmentation accuracy of these objects with multiple and variable scales in complex driving scenes. Second, a skip connection integrated with channel and spatial attention was proposed to focus the model on learning important travelability-related features while enabling efficient transmission from the encoder to the decoder features. The dual-attention skip connection can further enhance the detection performance on unstructured roads. An unstructured road driving dataset containing various real driving scenes of suburbs, villages, and campuses was constructed by several means. Experimental results on an unstructured road driving dataset demonstrate that the proposed model can effectively segment a real driving scene into a highly recommended, low-recommended, or non-recommended drivable region, and background. It benefits from the M-shaped deep architecture integrated with a multiscale interaction strategy and a dual attention mechanism. The proposed model achieves a mean intersection over union of 92.46% and an average detection speed of 22.7 frames per second, which has the advantages of both effectiveness and efficiency, and surpasses other existing mainstream models in the unstructured road drivable region detection task.