Dataset for a Unified Parallel Dual-Network Architecture for Furrow Following and Weed Detection Using Real Greenhouse Data

This work presents a unified vision-based perception framework for autonomous agricultural robots that combines CNN-FC-based furrow following with U-Net segmentation using two onboard RGB cameras. One camera feeds the command prediction network, whereas the second camera feeds the segmentation network. This approach addresses navigation and weed perception independently; both networks are trained in the same loop, where the proposed system integrates both tasks using separate models. Moreover, a low-cost robotic platform operating with real data is used for the experiments. Experimental results demonstrate strong generalization performance. Thus, the CNN-FC achieves a test mean absolute error of 2.34$^\circ$ for steering angle and 12.08 PWM units for velocity, corresponding to normalized errors of 0.058 and 0.080 of their respective operational ranges. Whereas, the U-Net achieves the following test values: pixel accuracy of 0.987, Dice of 0.916, IoU of 0.846, mIoU of 0.909, and mPA of 0.968. These results indicate that accurate and stable furrow-following control command prediction together with image segmentation can be achieved using a lightweight CNN--FC architecture combined with a U-Net architecture.