A Multispectral 3-D Vision System for Invertebrate Detection on Crops

Real-time detection and identification of invertebrates on crops is a necessary capability for integrated pest management, however, this challenging task has not been well-solved. Multispectral or hyperspectral machine vision systems have shown advantages for efficient and accurate detection and identification of certain invertebrate pests. However, only using spectral information has limited the capability for detection, especially for some camouflaged pests on host plants. Three-dimensional (3-D) object representations are being intensively studied for multiview object recognition and scene understanding in many fields. However, because of the lack of proper data collection methods and robust algorithms, 3-D technologies have not yet attained applications for detecting invertebrates. We have developed a multispectral 3-D vision system, which can create denser point clouds of plants and pests using the multispectral images of ultraviolet, blue, green, red, and near-infrared. An algorithm named local variance of normals was designed, which can distinguish broad leaves from relatively larger pests in noisy point clouds. The vision system could aid integrated pest management systems for pest monitoring, or could be used as a sensor of an automatic pesticide sprayer.