Agricultural robot plant automatic detection integrating visual navigation and phenotype recognition

With the continuous improvement of the intelligence level of facility agriculture, agricultural robots are undertaking more and more autonomous tasks in greenhouse environments, and the multifunctional integration of visual perception systems has become a key technological bottleneck. A perception system for agricultural robots that integrates visual navigation and phenotype recognition is developed to address issues such as path recognition being susceptible to environmental interference and poor real-time plant detection. The system consists of a path navigation module and a plant detection module. The former introduces an image segmentation method based on visual transformation structure to extract agricultural path information. The latter adopts a lightweight instance segmentation structure to achieve precise segmentation and structural localization of crop phenotype regions. In the navigation model test, in the rain and fog disturbance scene, the average delay is 51.0 ms, the frame rate is 44.2 FPS, and the control jitter amplitude is 1.36°. The test results of the detection module show that its boundary F1 values for Tomato, Cucumber, Pepper, and Lettuce crops are 90.2%, 87.6%, 88.8%, and 86.3%, respectively. The experimental results show that the proposed scheme reduces inference delay while ensuring accuracy, has good environmental adaptability and edge deployment potential, and demonstrates good robustness and practicality in complex greenhouse environments.