# “慧眼识株”–基于YOLO机器视觉农作物病虫害智能识别监测预警农保机器人

**Huiyan Shizhu | YOLO crop pest/disease inspection and early-warning agricultural-protection robot**

## Problem

Field pest/disease inspection needs mobility, edge vision, environmental sensing and remote supervision. This prototype combines all four in a low-cost agricultural protection robot.

## Architecture

Camera -> Raspberry Pi/Linux -> YOLOv5-Lite -> ONNX Runtime + OpenCV
STM32F103C8T6 <- sensors/actuators -> BC28 NB-IoT -> Alibaba Cloud IoT

## Runtime data path

1. The camera captures field imagery and the Raspberry Pi prepares frames for edge inference.
2. YOLOv5-Lite produces a pest/disease result locally; ONNX Runtime and OpenCV keep the deployment independent of a cloud inference round trip.
3. The STM32 node samples environmental sensors, applies device-state logic and drives the local actuators.
4. BC28 publishes telemetry and receives remote commands through Alibaba Cloud IoT for supervision outside the field.

## Implementation

1. YOLOv5 was slimmed by removing Focus and using YOLOv5-Lite. The exported ONNX model runs with ONNX Runtime and OpenCV on Raspberry Pi Linux.
2. STM32F103C8T6 owns sensor sampling, actuator drivers, state machine and serial debug. BC28 carries telemetry and cloud commands over NB-IoT.
3. The sensing set covers air temperature, humidity, pressure, soil moisture and illumination. The open-hub wheel improves mobility in soft fields.
4. I worked across the IoT link, embedded inference deployment, control integration and mechanical/kinematic validation.

## Validation focus

- Checked the exported model through Raspberry Pi Linux inference and OpenCV result rendering.
- Verified the serial/debug path between the STM32 control layer and the sensing/actuation modules.
- Used the architecture, detection and prototype images as evidence for the complete edge-to-cloud workflow.

## My role

- Built the IoT telemetry and remote-command integration path.
- Deployed and optimized the exported YOLO model for Raspberry Pi edge inference.
- Integrated the STM32 control layer with sensing, actuation and serial-debug workflows.
- Supported appearance/structure work, inverse kinematics and stability validation for the field robot.

## Evidence

![YOLO detection](yolo-detection-result.png)

![YOLO architecture](yolo-architecture.png)

![Field robot](field-robot.png)

## Deliverables

- [design-spec.docx](design-spec.docx): design specification, hardware/software/test sections and photo evidence.
- [presentation.pptx](presentation.pptx): project presentation.
- The source folder also contains patent notices and a demo video; large binaries are intentionally kept out of the profile repository.

## Outcome

Provincial second prize in Tianjin New Engineering Practice Innovation Competition; first prize in the university Electronic Design Innovation Practice Competition; utility model patent application in 2025.