节点文献

基于双目视觉的种植前期农田边界距离检测方法

Field Boundary Distance Detection Method in Early Stage of Planting Based on Binocular Vision

  • 推荐 CAJ下载
  • PDF下载
  • 不支持迅雷等下载工具,请取消加速工具后下载。

【作者】 洪梓嘉; 李彦明; 林洪振; 贡亮; 刘成良;

【Author】 HONG Zijia;LI Yanming;LIN Hongzhen;GONG Liang;LIU Chengliang;School of Mechanical Engineering, Shanghai Jiao Tong University;State Key Laboratory of Mechanical System and Vibration;

【通讯作者】 李彦明;

【机构】 上海交通大学机械与动力工程学院; 机械系统与振动国家重点实验室;

【摘要】 针对目前基于全球导航卫星系统技术的农机自动驾驶地头转弯方法的局限性,提出了基于双目视觉的农田田埂边界的识别和测距方法,对具体方法的可行性、适用性及约束条件进行了分析。针对光照变化大、重复纹理多的农田环境,双目立体匹配的代价计算步骤采用了Census变换和截断梯度融合的方法、代价聚合步骤采用了多尺度代价合并的分割树算法,可快速得到良好的视差图。针对农田地面不平坦及作物生长高度不均的实际情况,对视差图构建的三维点云进行了自适应阈值点云提取和干扰消除等操作,实现了田埂边界的识别。另外,根据农田信息,对计算的平均边界距离进行了校正。实验表明,此算法可以实现早期作业农田的边界距离检测,对前方5~10 m的田埂识别率达到99%,测距精度随着检测距离的减小而提高,5 m时的测距误差约0.075 m。在NVIDIA Jetson TX2硬件平台上,算法运行时间约0.8 s,对于行驶速度小于1.5 m/s的农机可满足作业的实时性要求。

【Abstract】 Aiming at the limitations of the current agricultural machinery automatic driving headland turning method based on global navigation satellite system technology, a binocular vision-based identification and ranging method of farmland ridge boundary was proposed, and the feasibility, applicability and constraints of the specific method were analyzed. In view of the farmland environment with large illumination changes and many repeated textures, Census transform and truncated gradient were integrated to calculate the cost of stereo matching, and cross-scale cost merging algorithm based on segment-tree was used in the cost aggregation step, which can quickly get a good parallax diagram. After constructing a three-dimensional point cloud from a parallax diagram, in view of the actual situation of uneven farmland ground and uneven crop growth height, the adaptive threshold point cloud extraction and interference elimination were carried out, so as to realize the recognition of field ridge boundary. In addition, according to the farmland information, the calculated average boundary distance was corrected. The experimental results showed that this algorithm can realize the boundary distance detection of the early working farmland, and the recognition rate of the algorithm can reach 99% for the ridge of 5~10 m in front of the field of view. The ranging accuracy was increased with the decrease of the detection distance, and the ranging error at 5 m was about 0.075 m. On NVIDIA Jetson TX2 hardware platform, the running time of the algorithm was about 0.8 s, which can meet the real-time requirements of the operation for the agricultural machinery with a driving speed less than 1.5 m/s.

【基金】 国家重点研发计划项目(2019YFB1312301)
  • 【文献出处】 农业机械学报 ,Transactions of the Chinese Society for Agricultural Machinery , 编辑部邮箱 ,2022年05期
  • 【分类号】TP391.41;S220.39
  • 【被引频次】2
  • 【下载频次】621
节点文献中: 

本文链接的文献网络图示:

本文的引文网络