Computer Vision Problems in 3D Plant Phenotyping

In recent years, there has been significant progress in Computer Vision based plant phenotyping (quantitative analysis of biological properties of plants) technologies. Traditional methods of plant phenotyping are destructive, manual and error prone. Due to non-invasiveness and non-contact properties as well as increased accuracy, imaging techniques are becoming state-of-the-art in plant phenotyping. Among several parameters of plant phenotyping, growth analysis is very important for biological inference. Automating the growth analysis can result in accelerating the throughput in crop production. This thesis contributes to the automation of plant growth analysis. First, we present a novel system for automated and non-invasive/non-contact plant growth measurement. We exploit the recent advancements of sophisticated robotic technologies and near infrared laser scanners to build a 3D imaging system and use state-of-the-art Computer Vision algorithms to fully automate growth measurement. We have set up a gantry robot system having 7 degrees of freedom hanging from the roof of a growth chamber. The payload is a range scanner, which can measure dense depth maps (raw 3D coordinate points in mm) on the surface of an object (the plant). The scanner can be moved around the plant to scan from different viewpoints by programming the robot with a specific trajectory. The sequence of overlapping images can be aligned to obtain a full 3D structure of the plant in raw point cloud format, which can be triangulated to obtain a smooth surface (triangular mesh), enclosing the original plant. We show the capability of the system to capture the well known diurnal pattern of plant growth computed from the surface area and volume of the plant meshes for a number of plant species. Second, we propose a technique to detect branch junctions in plant point cloud data. We demonstrate that using these junctions as feature points, the correspondence estimation can be formulated as a subgraph matching problem, and better matching results than state-of-the-art can be achieved. Also, this idea removes the requirement of a priori knowledge about rotational angles between adjacent scanning viewpoints imposed by the original registration algorithm for complex plant data. Before, this angle information had to be approximately known. Third, we present an algorithm to classify partially occluded leaves by their contours. In general, partial contour matching is a NP-hard problem. We propose a suboptimal matching solution and show that our method outperforms state-of-the-art on 3 public leaf datasets. We anticipate using this algorithm to track growing segmented leaves in our plant range data, even when a leaf becomes partially occluded by other plant matter over time. Finally, we perform some experiments to demonstrate the capability and limitations of the system and highlight the future research directions for Computer Vision based plant phenotyping.