Online cooperative perception of working width and navigation lines for full-width rice and wheat harvesting robots

To address the problems of missed or incomplete harvesting and non-full-width harvesting caused by header offset during the autonomous operation of rice and wheat harvesting robots, accurate real-time estimation of the working width and navigation path is essential. However, existing methods are limited by the high computational complexity of semantic segmentation models, reliance on expensive radar or depth cameras for working width detection, and insufficient accuracy and robustness of navigation line extraction in complex scenarios such as headland turns. To overcome these limitations, this study proposes an online collaborative detection method that integrates lightweight semantic segmentation with keypoint-based geometric computation to estimate cutting width and navigation paths. First, a field operation image dataset covering multiple machine types and operating conditions was constructed. Second, the DeepLabv3+ model was enhanced by incorporating a lightweight MobileViT backbone, a CBAM attention mechanism, and depthwise separable dilated convolution, resulting in a more compact model with improved representation of the operating region. Based on this, a low-cost cutting width detection method using key pixel coordinate mapping was designed, eliminating the need for depth sensors through geometric computations on boundary keypoints extracted from multiple regions. Furthermore, navigation lines were extracted using boundary constraints and weighted fitting, improving robustness in complex scenarios, such as turning operations. The improved segmentation model achieved an mIoU and mPA of 91.62% and 94.80%, respectively, on the test set, while requiring only 2.62% of the parameters of the original model. The average absolute error (MAE) of cutting width detection was below 0.10 m, and the average relative error (MRE) was below 3.00%. Moreover, the average angular deviation in navigation line extraction was below 3°, and the lateral error was less than 6 pixels, demonstrating significant advantages over traditional methods in headland turning scenarios. By coordinating perception and correction using swath width and navigation lines, the effective working width utilization rate increased by approximately 35%. This method demonstrated robustness under varying illumination, dust, and travel speed conditions, providing a high-precision, low-cost visual perception and autonomous operation solution for rice and wheat harvesting robots.