Electronic Thesis and Dissertation Repository


Master of Science


Electrical and Computer Engineering


Dr. Kenneth A.McIsaac


Image classification and segmentation are the two main important parts in the 3D vision system of a harvesting robot. Regarding the first part, the vision system aids in the real time identification of contaminated areas of the farm based on the damage identified using the robot’s camera. To solve the problem of identification, a fast and non-destructive method, Support Vector Machine (SVM), is applied to improve the recognition accuracy and efficiency of the robot. Initially, a median filter is applied to remove the inherent noise in the colored image. SIFT features of the image are then extracted and computed forming a vector, which is then quantized into visual words. Finally, the histogram of the frequency of each element in the visual vocabulary is created and fed into an SVM classifier, which categorizes the mushrooms as either class one or class two. Our preliminary results for image classification were promising and the experiments carried out on the data set highlight fast computation time and a high rate of accuracy, reaching over 90% using this method, which can be employed in real life scenario.

As pertains to image Segmentation on the other hand, the vision system aids in real time identification of mushrooms but a stiff challenge is encountered in robot vision as the irregularly spaced mushrooms of uneven sizes often occlude each other due to the nature of mushroom growth in the growing environment. We address the issue of mushroom segmentation by following a multi-step process; the images are first segmented in HSV color space to locate the area of interest and then both the image gradient information from the area of interest and Hough transform methods are used to locate the center position and perimeter of each individual mushroom in XY plane. Afterwards, the depth map information given by Microsoft Kinect is employed to estimate the Z- depth of each individual mushroom, which is then being used to measure the distance between the robot end effector and center coordinate of each individual mushroom. We tested this algorithm under various environmental conditions and our segmentation results indicate this method provides sufficient computational speed and accuracy.