Estimating apple yield using video images and deep learning
LE3 .A278 2021
2021
Silver, Danny
Acadia University
Master of Science
Masters
Computer Science
Our research uses low-cost video, computational vision techniques, and deep learning methodologies to estimate apple yield while the apples are still on the tree. We captured a video of the Sweet Tango apple variety at Pomona Farms, Canard, Nova Scotia from both sides of an orchard row. A convolutional neural network (CNN) based on the YOLOV3 architecture is trained to detect apples in the sequence of video images. The tracking of each detected apple is done using a Kalman filter. The Hungarian algorithm, inclusive of motion and appearance information is used to solve the association problem between the previously tracked and the newly detected apples. The count is derived using this tracking information, and the average size is determined by using each apple’s bounding box information from the detection module. The experiments show that the proposed framework performs best using daylight video images combined with transfer learning and data augmentation techniques. The final predictive model for apple detection has an AP (average precision) score of 90.72, the tracking model has an MT (mostly tracked) score of 36.09, and the system counts apples with an F1-score of 94.97%. We generate a video that shows the distinction between the apples that are being tracked to those that are counted by a change in the color of the bounding box surrounding the apple. Additionally, we developed proof-of-concept methods to (1) avoid double counting from videos captured on both sides of the row and to (2) better estimate the size of each apple.
The author retains copyright in this thesis. Any substantial copying or any other actions that exceed fair dealing or other exceptions in the Copyright Act require the permission of the author.
https://scholar.acadiau.ca/islandora/object/theses:3697