A comparative study of fruit detection and counting methods for yield mapping in apple orchards

Abstract

We present a modular end‐to‐end system for yield estimation in apple orchards. Our goal is to identify fruit detection and counting methods with the best performance for this task. We propose a novel semantic segmentation‐based approach for fruit detection and counting and perform extensive comparative analysis against other state‐of‐the‐art techniques. This is the first work comparing multiple fruit detection and counting methods head‐to‐head on the same data sets. Fruit detection results indicate that the semisupervised method, based on Gaussian Mixture Models, outperforms the deep learning‐based methods in the majority of the data sets. For fruit counting though, the deep learning‐based approach performs better for all of the data sets. Combining these two methods, we achieve yield estimation accuracies ranging from 95.56% to 97.83%.

MinneApple: A Benchmark Dataset for Apple Detection and Segmentation

“…Similarly, researchers released specialized datasets for autonomous driving [3], [30], [10] or pedestrian detection [8], [5]. While precision automation and automated yield mapping have seen much research effort [1], [29], [25], [2], [13], [12], each of these papers used their own datasets of varying completeness and level of detail.…”

“…Gongal et al [11] offer a comprehensive overview of these early detection methods. More recent papers used object detection networks to detect fruits [28], [1], [12]. Sa et al [28] used a combination of NRI and RGB images of fruits in indoor environments.…”

“…They split the dataset of 71 images 50/50 between training and testing. In our previous work [25], [26], [27], [12] we presented results on HD sized images, showing parts of an orchard row. We presented multiple methods including semi-supervised Gaussian Mixture Model (GMM) a Faster R-CNN object detector and a semantic segmentation network.…”

“…The network was initialized with weights pretrained on COCO [19]. We follow [1], [12] in our choice of parameters for optimization.…”

“…UNet (not pretrained): A semantic segmentation network, based on a fully convolutional network architecture [23], [12]. The images in the train and test sets are split into 224 × 224 sized chunks, and the weights of the network are initialized randomly.…”