3D machine vision system for robotic weeding and plant phenotyping
The need for chemical free food is increasing and so is the demand for a larger supply to feed the growing global population. An autonomous weeding system should be capable of differentiating crop plants and weeds to avoid contaminating crops with herbicide or damaging them with mechanical tools. For the plant genetics industry, automated high-throughput phenotyping technology is critical to profiling seedlings at a large scale to facilitate genomic research. This research applied 2D and 3D imaging techniques to develop an innovative crop plant recognition system and a 3D holographic plant phenotyping system.
A 3D time-of-flight (ToF) camera was used to develop a crop plant recognition system for broccoli and soybean plants. The developed system overcame the previously unsolved problems caused by occluded canopy and illumination variation. Both 2D and 3D features were extracted and utilized for the plant recognition task. Broccoli and soybean recognition algorithms were developed based on the characteristics of the plants. At field experiments, detection rates of over 88.3% and 91.2% were achieved for broccoli and soybean plants, respectively. The detection algorithm also reached a speed over 30 frame per second (fps), making it applicable for robotic weeding operations.
Apart from applying 3D vision for plant recognition, a 3D reconstruction based phenotyping system was also developed for holographic 3D reconstruction and physical trait parameter estimation for corn plants. In this application, precise alignment of multiple 3D views is critical to the 3D reconstruction of a plant. Previously published research highlighted the need for high-throughput, high-accuracy, and low-cost 3D phenotyping systems capable of holographic plant reconstruction and plant morphology related trait characterization. This research contributed to the realization of such a system by integrating a low-cost 2D camera, a low-cost 3D ToF camera, and a chessboard-pattern beacon array to track the 3D camera's position and attitude, thus accomplishing precise 3D point cloud registration from multiple views. Specifically, algorithms of beacon target detection, camera pose tracking, and spatial relationship calibration between 2D and 3D cameras were developed. The phenotypic data obtained by this novel 3D reconstruction based phenotyping system were validated by the experimental data generated by the instrument and manual measurements, showing that the system has achieved measurement accuracy of more than 90% for most cases under an average of less than five seconds processing time per plant.