Reconstructing maize plants at the seedling stage using the improved ORB-SLAM2 algorithm

Three-dimensional (3D) reconstruction for crops is often limited to a prolonged duration and suboptimal quality. In this study, an improved framework was proposed as the Oriented FAST and Rotated BRIEF-SLAM2 (ORB-SLAM2) system, in order to specifically reconstruct the seedling-stage maize plants. Several advanced computer vision techniques were integrated to achieve a robust and efficient reconstruction. Initially, the Fast Library for Approximate Nearest Neighbors (FLANN) algorithm was employed to combine with the Random Sample Consensus (RANSAC) algorithm, in order to perform high-accuracy feature matching on the multi-view images of the maize seedlings. This hybrid matching strategy was effectively identified to align the keypoints over the different images. While the outliers and incorrect matches were filtered out for the subsequent 3D processing. Feature matching and sparse point cloud generation were then realized using the modified ORB-SLAM2 front-end. A Multi-View Stereo (MVS) algorithm was seamlessly integrated into the pipeline. This MVS module was utilized for the pose estimations using a precise camera. The feature was performed on dense matching and depth estimation. Ultimately, a dense and accurate 3D point cloud model was generated to represent the morphology and spatial structure of the target maize plants. After that, the key architectural parameters were extracted from two maize cultivars, such as the plant height, leaf length, and stem diameter. A comparison was then made between the conventional and manual measurements in order to evaluate the practical utility and accuracy of the reconstructed models. Experimental results demonstrate that the superior performance of the integrated approach was achieved in the accuracy, reliability, and overall effectiveness of the non-contact phenotype using imaging. Specifically, the FLANN+RANSAC combination achieved a high correctness rate of 89.00% in the feature matching. The dense reconstruction produced the point clouds with an average density of 7.13×10⁵ points per model. Remarkably, the entire reconstruction, from the image input to dense point cloud output, was required for an average time of only 15.32 min, indicating its significant efficiency. All architectural parameters were extracted from the 3D models for the seedling maize plants. The errors were consistently maintained within a 10% threshold. There was a significant correlation between the extraction and the manual measurements, indicating the strong robustness. In conclusion, the high speed and precision were realized in plant phenotyping. The computational time was substantially reduced without compromising the reconstruction quality, thus offering a rapid, accurate, and non-destructive solution. Therefore, this work can provide a solid theoretical foundation and a practical, high-performance technical framework for the automatic acquisition of the 3D architectural traits in the maize plants during the critical seedling stage. The considerable potential was provided to advance the high-throughput plant phenotyping. The efficient morphological analysis can also accelerate crop genetics and breeding in precision agriculture.