TREEREG
Detection-based Registration

Overview

Aerobotics is a company in the Precision Agriculture sector. A key functionality of their product is the ability to provide trend analysis of the performance of an orchard on an individual tree level over time. In order to enable this, successive orthomosaic images of the same orchard at different dates have to be accurately registered. Image Registration is the process of aligning different images of the same scene taken at different times, angles and sensors. Aerobotics have developed a feature extractor to aid this process, a tree instance segmentation model that produces a boundling polygon over individual tree detections.

The main aim of this project is to use these tree detections to register the orthomosaic images of orchards at different dates. There are two main challenges to this project: The registration method needs to be robust to the temporally unstable nature of trees (they can grow, whither, shift or be destroyed). The method also needs to be robust to the local deformations of each image as a result of the orthomosaic generation process.

Two Point Cloud Registration (PCR) algorithms were implemented as potential solutions to this problem: Coherent Point Drift (CPD) and Robust Point Matching (RPM). Both methods were shown to produce an accurate registration of the input tree detections modelled as point clouds (under certain simulated deformations). 

The above figure displays the overall process of this project. Two input orthomosaic images and their associated tree detection polygons are provided. From this input point clouds are produced by taking the centroids of each polygon. These point clouds are then aligned using a PCR algorithm. The implementation and results of each PCR algorithm are outlined in detail on their individual pages.