Abstract: In order to realize the fast and accurate identification of the roadable area of unstructured forest, an unstructured road segmentation model based on improved DeepLabV3+ was proposed to solve the problems such as unclear boundary, irregular road shape and road coverage. By using MobileNetV3 network instead of the traditional DeepLabV3+ backbone network to achieve lightweight design, the network image segmentation speed and real‑time significantly was improved. CBAM attention mechanism was introduced in the decoder part of backbone network, and by adjusting the parameters of ASPP module, the feature extraction and recognition of unstructured road in the boundary region were enhanced. The fusion loss function was used to improve the convergence rate and accuracy of the model and avoid the error detection area in the complex environment. The results show that the average detection frame number of DeepLabV3+ improved by 26.69 frames/s, the detection rate increased by about 54% compared with the original model, and the detection accuracy increased to 91.26%. At the same time, no missing detection, false detection and unclear boundary segmentation occurred under various conditions such as strong light, backlight and road water, which can provide technical reference for unstructured road automatic driving.

Key words: unstructured roads, semantic segmentation, DeepLabV3+, attention mechanism, loss function

摘要： 为实现非结构化林间道路可行使区域的快速准确识别，针对林间道路边界不明显、道路形状不规范以及道路覆盖等问题，提出一种基于改进DeepLabV3+的林地非结构化道路分割模型。使用MobileNetV3网络代替传统DeepLabV3+主干网络以实现轻量化设计，使图像分割速度及实时性显著提升；在主干网络解码器部分引入CBAM注意力机制，通过对ASPP模块参数调整，增强对非结构化道路在边界区域的特征提取与识别；采用融合损失函数，提高模型收敛速率及准确度，避免模型在复杂环境下出现错误检测区域。结果表明，改进后的DeepLabV3+检测平均帧数提升26.69帧/s，较原模型检测速率提升约54%，检测准确率提升至91.26%，同时，在强光、逆光以及路面积水等多种情况下均未出现漏检、误检和边界分割不清晰等现象，为非结构化道路自动驾驶提供技术参考。

关键词: 非结构化道路, 语义分割, DeepLabV3+, 注意力机制, 损失函数