Botian needed a more robust way to guide agricultural robots in rice fields, because traditional image-processing methods could not reliably extract seedling center lines under changing light, floating weeds, missing seedlings, and other noise in unstructured field environments.
Reported impact
Botian improved the reliability and speed of autonomous field navigation, reducing dependence on manual labor and making agricultural robots more practical in complex rice-field environments.
Botian built an automatic paddy-field navigation-line detection system on PaddlePaddle and PaddleSeg so its agricultural robots could navigate more accurately in real field conditions. The workflow starts by collecting labeled images from two representative growth stages — one week and three weeks after transplanting — at different times of day and exposure settings so the training set covers changing brightness conditions. Botian then chose the ICNet semantic segmentation model because it offers a strong compromise between accuracy and real-time performance for embedded-style machine-vision use cases. Training runs on 640x512 inputs use configurable optimizers, augmentation strategies, and validation checkpoints, while VisualDL is used to inspect loss curves and pick the best model without overfitting. Inference segments the seedlings from background clutter and then extracts row center lines for robot guidance. This turned plant-row detection from a fragile hand-tuned image pipeline into a trainable AI perception stack suited to real agricultural automation.
Field images captured under different times and exposure settings are labeled and used to train ICNet through PaddleSeg on PaddlePaddle, the model segments seedlings from the noisy background, and the extracted row center lines are fed into agricultural robot navigation workflows.
Average angle error dropped from 3.27° to 1.65°. (original: 在试验图像中的平均角度误差从3.27°降低到了1.65°。)
Botian developed the paddy-field navigation-line detection system on PaddlePaddle and used the ICNet model in PaddleSeg to segment seedlings from the background and extract center lines precisely. (original: 基于百度飞桨(PaddlePaddle)深度学习平台研发了水田导航线自动检测系统。他们应用飞桨图像分割开发套件PaddleSeg中的ICNet模型将秧苗按列从背景中分割出来，并以此为基础实现了秧苗列中心线的精准提取。)
ICNet was selected because it achieves real-time performance with much lower memory and prediction time while keeping accuracy at an acceptable level. (original: ICNet...属于参数量较小的轻量级语义分割网络...在精度降低较少的情况下大幅度减少了预测时间和占用内存。)
The dataset included 400 images from each of two growth stages at different times of day and exposure settings. (original: 技术人员分别在一天内不同时间段采集插秧1周后和插秧3周后两个典型时期的水稻秧苗的图片各400幅，并且在每个时段通过设置不同的曝光时间以采集不同亮度的图片。)
The final system achieved over 95% extraction accuracy and about 300 ms processing time per image frame. (original: 提取准确率达到95％以上，处理每帧图像耗费的时间...达到300ms左右。)
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
Baidu AI Cloud Sample Apr 9, 2026 China Baowu Steel Group Baowu needed to move ladle status data, image data and production-process data safely to the cloud so it could detect hidden risks, optimize operations and lower storage costs. BML Full-Function AI Development Platform · Big Data Platform · IoT Core Baidu AI Cloud Sample Apr 8, 2026 SANY Heavy Machinery Co., Ltd. (三一重机) SANY needed to uphold its 'solve within one day' service standard for excavators and other machinery even when remote field engineers lacked deep experience and half of faults could not be diagnosed accurately without… Before the project, as many as 50% of problems could not be diagnosed accurately. Industrial AR Remote Assistance System · Intelligent Industrial Quality Inspection · Tiangong Edge-Cloud Integrated IoT Platform Baidu AI Cloud Sample Dec 26, 2025 China Iron and Steel Research Institute Group (中国钢研科技集团) China Iron and Steel Research needed a practical way to apply large-model and visual AI capabilities across metallurgical research and manufacturing, especially for metallography analysis, surface defect detection, and… 100 ms response time Baige AI Heterogeneous Computing Platform · Kunlun · Qianfan Large Model Platform 