Counting of Strawberries and Flowers in Fields using YOLOv4 and SORT

Abstract. Strawberry harvesting takes a lot of planning and labor. The fruits once ripe must be harvested as they do not last for long if left on the field. So, counting the flowers and strawberries, and predicting the yield would not only help in estimating the best time to harvest but also help in allocating resources efficiently, especially labor which is not easy to recruit. In this study, a 3-stage method is proposed to detect and count strawberries and flowers. The YOLO network was trained to localize and classify flowers and strawberries on the field. The strawberries can be harvested when completely red and can even be harvested when they are more than a quarter red depending on the situation. Hence, a customizable segregation algorithm is proposed which can help by giving the count of ripe, unripe strawberries for different thresholds according to the redness without retraining the YOLO model. The algorithm resulted in F1 scores of 0.956, 0.931, 0.918 for flower, ripe, unripe classes respectively on the video with similar conditions and 0.864, 0.858, 0.766 for flower, ripe, unripe classes respectively on the video with conditions different from the training data. The strawberries and flowers were tracked using multiple object tracker SORT and counted using a line cross algorithm. This system can help the user to predict and plan the growth and harvest of the strawberry plants.