Apple Detection and Counting Using Neural Networks

It is important to estimate the number of fruits which are going to be harvested, since that information allows farmers to make key decisions regarding the production, distribution and commercialization of crops. In this project, solutions for detecting and counting apples are explored, by using detection and tracking algorithms based on neural networks. Four neural network object detection models were trained: YOLOv5, YOLOv8 and two variants of Faster R-CNN, using public as well as self-made datasets. Three tracking algorithms (StrongSort, ByteTrack and OCSort) were tested using the trained models and labeled video datasets. These datasets were developed with recordings of a real-world orchard and recordings generated with a ROS-based apple field simulator, developed as part of this project. Finally, two fitting models based on linear regression were evaluated in order to minimize prediction errors. As a conclusion of this research, it was found that the YOLOv5 and YOLOv8 performances were better than the other two, achieving a mAP50 of 0.87 and 0.85. In terms of apple tracking, the performance of OCSort stands out, reaching a HOTA metric of 60.24% and running times in the order of 30 ms.