基于轻量化分割与热图回归的农业场景路径提取方法

Agricultural Scene Navigation Path Extraction Method Based on Lightweight Semantic Segmentation and Heatmap Regression

摘要:

为解决农业非结构化环境下机器人自主导航路径提取过程中路径断裂、中心线偏移、实时性差以及复杂环境适应性弱等问题,本文提出了一种基于轻量化分割与热图回归的农业场景导航路径提取方法。采用改进的Deeplabv3+语义分割模型,结合轻量化MobileNetV3作为主干特征提取网络,有效提升了模型在嵌入式设备中推理速度,同时保持了较高的分割精度。基于语义分割获得的可行驶区域掩码图,设计了一种Mobile-Unet轻量化热图回归网络,通过加权均方误差与感知损失联合的组合损失函数,增强了模型对热图高响应区域关注能力,并提高了热图预测结果细节还原能力与结构一致性。提出一种基于热值引导贪心搜索与非极大值抑制相结合的导航路径提取算法,能够在复杂农业场景下从热图中高效、稳定提取连续平滑的导航路径。在自建的农业场景数据集上开展试验,验证了语义分割模型、热图回归网络及路径提取算法性能。结果表明,本文方法在保证语义分割精度基础上,提升了模型在Jetson TX2嵌入式平台上的推理速度;在热图回归任务中,所设计的Mobile-Unet模型结合Mask输入方案获得了最低均方误差（MSE）与最佳结构相似性指标（SSIM）,且满足实时推理需求。所提方法兼顾了精度、速度与实际部署能力,具备良好的农业机器人导航应用前景。

Abstract:

Aiming to solve the problems of path discontinuity, centerline deviation, poor real-time performance, and weak adaptability to complex environments in unstructured agricultural environments of robot navigation path extraction, a navigation path extraction method for agricultural scenes was proposed based on semantic segmentation and heatmap regression. Firstly, an improved Deeplabv3+ semantic segmentation model was constructed by integrating the lightweight MobileNetV3 as the backbone network, which effectively improved the inference speed of the model in embedded devices while maintaining a high segmentation accuracy. Secondly, based on the drivable area mask obtained by semantic segmentation, a Mobile-Unet lightweight heatmap regression network was designed. Through the combined loss function of weighted mean square error and perceptual loss, the model's ability to focus on high-response areas of the heatmap was enhanced, and the detail restoration and structural consistency of the heatmap prediction results were improved. Finally, a navigation path extraction algorithm combining heat value-guided greedy search and non-maximum suppression (NMS) was proposed to efficiently and stably extract continuous and smooth navigation paths from the heatmaps in complex agricultural environments. Experiment was carried out on a self-built agricultural scene dataset, which verified the performance of the semantic segmentation model, heatmap regression network and path extraction algorithm. Results demonstrated that the proposed method maintained high segmentation accuracy while achieving superior inference speed on the Jetson TX2 embedded platform. In heatmap regression tasks, the proposed Mobile-Unet combined with Mask input achieved the lowest mean square error (MSE) and the highest structural similarity (SSIM), meeting the requirements for real-time navigation. Overall, the proposed method took into account accuracy, speed and practical deployment capabilities, and had good application prospects in agricultural robot navigation.