Automatic pear and apple detection by videos using deep learning and a Kalman filter

Abstract

Pears and apples in videos recorded while walking were detected automatically using a deep-learning-based method referred to as YOLO. The same fruits in the successive video frames were then identified using a Kalman filter. The average precision of the pear detection was 0.97, while the number of correctly counted pears was 226, out of 234. A YOLO v2 network with a larger input image size and data augmentation method contributed to the high accuracy in the counting. The pears and apples in the videos were counted automatically, within an absolute error of 10% under unstable light conditions and with greenish fruits.