Cotton Crop Inter-row Navigation Path Recognition Method Based on an Improved UNet Model

To tackle the problems of low accuracy, poor real-time performance, and the difficulty of deploying complex models for cotton row navigation path recognition, this paper proposes the UNet-Mv3L+ model, which integrates the advantages of the UNet and MobileNet-v3 Large models. The Labelme tool is used to annotate road information in collected images and create a dataset of cotton crop inter-row. MobileNet-v3 Large replaces the backbone feature extraction network of UNet, and a Convolutional Block Attention Module (CBAM) is added before each layer of feature map fusion to further enhance feature extraction performance from both channel and spatial perspectives. Based on the segmented regions extracted by this model, a spline curve fitting is used to generate navigation lines between cotton rows. The model training and evaluation were completed on an augmented cotton crop inter-row training dataset. Experimental results show that the improved UNet-Mv3L+ model achieves an Mean Intersection over Union (MIoU) and Mean Pixel Accuracy (MPA) of 96.85% and 98.47%, respectively, under various lighting conditions, demonstrating stable pixel-level segmentation of cotton rows. Compared to the original UNet model, the segmentation accuracy has improved by 1.42% in MIoU and 2.34% in MPA. Additionally, navigation path recognition accuracy tests show an average pixel error of 9.2 pixels and an average distance error of 0.025m. Given that the width between cotton rows is 0.78m, the average distance deviation accounts for 3.2%. The model's theoretical inference speed for processing RGB image segmentation regions can reach 22.15 frames per second. The research method presented in this paper provides an effective reference for solving the visual navigation task of cotton crop inter-row.