Kiwifruit Counting Using Kiwidetector and Kiwitracker

Efficient fruit detection and counting are crucial to improve fruit industrial efficiency and assist the farmers to develop reasonable harvesting strategies in advance while significantly reducing human labors and wastage. In this paper, digital video and image datasets of kiwifruits are collected to train a deep learning-based fruit counting model. The model consists of two molecular algorithms: KiwiDetector for kiwifruit detection and KiwiTracker for kiwifruit tracking and counting. The KiwiDetector algorithm is based on the state-of-the-art YOLOv7 algorithm. The KiwiTracker algorithm is from Kalman filtering and global matching for target tracking. The KiwiDetector module obtained mAP 0.937 after model training based on our collected kiwifruit dataset. The KiwiTracker module has an average counting precision 0.802 after model testing based on different videos containing kiwifruits. The methods can assist us in estimating yields efficiently and provide technical references for agricultural automation.