Division of Information Technology, Engineering and the Environment
School of Advanced Manufacturing and Mechanical Engineering
School of Advanced Manufacturing and Mechanical Engineering
Thesis Abstract
In horticultural industry, conventional harvesting is done by ‘handpicking’ methods to remove hundreds of fruits such as citrus fruits in random spatial locations on the individual fruit trees. It is well known that harvesting fruits in a large scale is still inefficient and not cost effective. To solve this challenging task, mechanical harvesting systems have been investigated and practiced to enhance profitability and efficiency of horticultural businesses. However they often damage fruits in the harvesting process. Development of efficient fruit removal methods are required to maintain the fruits quality. The difficulties are multi-factor by the nature of the data, the limitation of the sensor and an unstructured dynamic grove. This project focuses on two parts. The first part is a practical extension on citrus fruit tree imaging and measurement using multi-spectral imaging method. The second part is the study on the citrus image data and the methods behind those data.
To acquire and practice the multi-spectral imaging, two issues needs to be addressed such as the registration on the images from multi sensor and the fusion methods. The bi-camera and cold mirror acquisition scheme has been proposed to extend a citrus imaging and measurement in a dynamic environment. By our proposed cold mirror system, the registration could be bypassed and guaranteed by the software based calibration on both cameras on the cold mirror fixture in a kinematic alignment to certain precision. Therefore the serious issue such as the local distortion or occlusion can be improved significantly. The fusion of two components from visible area and near infrared area is combined on the lower filtered components by wavelets decomposition such as Ingrid Daubecies wavelets.
On top of the data, the centrepiece is the study on the multispectral data in different wavelength band from a bi-camera and cold mirror system for the data learning. Data collection to be studied includes the original visible data and the fused composition data by fusing the visible and near infrared data, and the physical optic modified data. Study on the citrus data will cover main methods from the literature including the color indices, linear discriminant analysis such as statistical Fisher linear discriminant to the hyperplane by using such as single perceptron or multilayer perceptron architecture, and the precise method using support vector machine. The algorithm for the real time application based on the preliminary study considers the prediction of the cluster of the citrus and the incremental movement for the distance estimation for the tool coordinates position of the robot in a dynamic unstructured environment.
