作物行识别算法的虚拟试验方法

Virtual Test Method for Algorithm of Crop Row Detection

针对作物行识别算法的传统开发过程对田间作物生长周期依赖性较强,错过适当的田间图像采集时期将直接导致算法开发周期延长等问题,提出一种基于虚拟场景的作物行识别算法测试方法,即在虚拟环境下模拟农田作物行场景和图像采集系统,运用虚拟作物行图像测试作物行的识别算法。该方法在虚拟现实环境下建立作物行场景模型;提出一种融合建模法,根据作物和杂草的几何特征建立对应的三维几何模型;根据实际田间作物的空间分布特征,建立株距、行距可调的田间作物行模型;以Vega Prime为视景仿真工具,通过配置投影模式、渲染模式、视点位姿和图像采集规格,构建图像采集系统,输出作物行场景图像。以苗期棉花作物行为建模对象,对一种经过田间试验验证的双目视觉作物行识别算法进行测试试验。对比实际棉田图像对应的试验结果,同一作物行识别算法的识别正确率、偏差角和图像处理时间均相近。结果表明,本文建立的虚拟棉田作物行与实际棉田作物行场景相近,能够用于作物行识别算法的测试。

Crop row detection is an intrinsic issue for machine vision-based guidance of agricultural machinery. The classical development for algorithm of crop row detection is based on real field images.Real field image acquisition is related to crop growth cycle closely,which is greatly affected by local district,climate and crop growth status. If the appropriate period of real field image acquisition was missed,the development for algorithm of crop row detection would be delayed directly and the cost would also be increased. In order to improve the efficiency of development of crop row detection and save cost,a new method based on virtual reality to test the crop row detection was proposed. Crop rows were simulated in virtual test environment to provide image data for the development of crop row detection. The proposed method consisted of two parts which were simulation of crop row field and virtual image acquisition. The 3 DS Max and Multigen-Creator were used to build models. The Vega Prime was used to simulate the models in virtual environment. To simulate the real crop row field, the individual characteristics and group characteristics were considered during the modeling, respectively. The simulation of a virtual crop row field was composed of the modeling of single crop and weed. A fusion method was proposed to build models of the single crop and weed. Specifically,the leaf of crop and weed whose spatial feature were anisotropy was modeled with the counterdraw method; the stem and petiole of crop and weed whose feature were concealed by leaves were modeled with the billboard method or cross method. To express the group characteristics of real crop row field,a parametric modeling method was proposed based on random sampling. The sample libraries of crop and weed were composed by several models,respectively. Every crop and weed was placed in the virtual environment at random position and rotate angle within specific thresholds. The spaces in row and column directions were set according to thereal field. In order to acquire the virtual field image,an image acquisition system was designed in Vega Prime. The asymmetric projection was used to simulate the lens of a real camera. The number of render window was equal to the number of lens. The attitude of viewpoint simulated the relative attitude between camera and vehicle. The specification of viewfinder was adjusted to acquire images with different sizes.After the image of virtual field was acquired,the algorithm of crop row detection could be tested. Crop rows of cotton in seeding stage were simulated as an example. A binocular vision-based algorithm of crop row detection was tested with the acquired images of virtual cotton field. The experimental results of virtual field and real field were similar. Results showed that the proposed modeling method can build the virtual cotton field conveniently,which provided sufficient images for testing the algorithm of crop row detection.