Design of a Dynamic Obstacle Avoidance System for Greenhouse Robots Based on the Fusion of YOLOv5 and Ultrasonic Sensing

To address the limitations of single-sensor obstacle avoidance in greenhouse environments characterized by complex lighting and diverse obstacle morphologies, this study proposes and implements a dynamic obstacle avoidance system based on the fusion of visual and ultrasonic multi-sensor information. The system employs an STM32F407 microcontroller as the real-time control core and a Canaan K230 edge AI processor as the dedicated visual processing unit, constructing a hardware platform that integrates a lightweight YOLOv5s model, multiple ultrasonic sensors, and a dual-track drive module. The model is trained and converted using "VSCode + PyTorch," while embedded software programming is accomplished with "Keil MDK + CanMV IDE," enabling a complete implementation from algorithm to hardware. This paper achieves a complementary advantage of long-range observation and short-range avoidance by fusing semantic recognition results with rangig data in real-time on the STM32. Experimental validation in a simulated greenhouse environment demonstrate that the proposed fusion system attains an obstacle avoidance success rate of 94%, with an average decision response time of 305 milliseconds, representing a significant performance enhancement over single-sensor solutions. This research provides a low-cost, robust, and practical solution for autonomous navigation of mobile robots in greenhouses.