Automatic plant features recognition using stereo vision for crop monitoring

Machine vision and robotic technologies have potential to accurately monitor plant parameters which reflect plant stress and water requirements, for use in farm management decisions. However, autonomous identification of individual plant leaves on a growing plant under natural conditions is a challenging task for vision-guided agricultural robots, due to the complexity of data relating to various stage of growth and ambient environmental conditions. There are numerous machine vision studies that are concerned with describing the shape of leaves that are individually-presented to a camera. The purpose of these studies is to identify plant species, or for the autonomous detection of multiple leaves from small seedlings under greenhouse conditions. Machine vision-based detection of individual leaves and challenges presented by overlapping leaves on a developed plant canopy using depth perception properties under natural outdoor conditions is yet to be reported. Stereo vision has recently emerged for use in a variety of agricultural applications and is expected to provide an accurate method for plant segmentation and identification which can benefit from depth properties and robustness. This thesis presents a plant leaf extraction algorithm using a stereo vision sensor. This algorithm is used on multiple leaf segmentation and overlapping leaves separation using a combination of image features, specifically colour, shape and depth. The separation between the connected and the overlapping leaves relies on the measurement of the discontinuity in depth gradient for the disparity maps. Two techniques have been developed to implement this task based on global and local measurement. A geometrical plane from each segmented leaf can be extracted and used to parameterise a 3D model of the plant image and to measure the inclination angle of each individual leaf. The stem and branch segmentation and counting method was developed based on the vesselness measure and Hough transform technique. Furthermore, a method for reconstructing the segmented parts of hibiscus plants is presented and a 2.5D model is generated for the plant. Experimental tests were conducted with two different selected plants: cotton of different sizes, and hibiscus, in an outdoor environment under varying light conditions. The proposed algorithm was evaluated using 272 cotton and hibiscus plant images. The results show an observed enhancement in leaf detection when utilising depth features, where many leaves in various positions and shapes (single, touching and overlapping) were detected successfully. Depth properties were more effective in separating between occluded and overlapping leaves with a high separation rate of 84% and these can be detected automatically without adding any artificial tags on the leaf boundaries. The results exhibit an acceptable segmentation rate of 78% for individual plant leaves thereby differentiating the leaves from their complex backgrounds and from each other. The results present almost identical performance for both species under various lighting and environmental conditions. For the stem and branch detection algorithm, experimental tests were conducted on 64 colour images of both species under different environmental conditions. The results show higher stem and branch segmentation rates for hibiscus indoor images (82%) compared to hibiscus outdoor images (49.5%) and cotton images (21%). The segmentation and counting of plant features could provide accurate estimation about plant growth parameters which can be beneficial for many agricultural tasks and applications