Research On Apple Yield Measurement Model Based On YOLO And DeepSort Vision Algorithm

Apples are rich in nutritional value and are one of the main daily edible fruits for the masses,with a great demand in the market.Red Fuji apples are commonly grown in Zhaotong,Yunnan,and are an irreplaceable pillar of the local agricultural economy.At present,apple production is mainly measured by hand,which is not only time-consuming but also inefficient.With the increasing degree of automation in orchards,in order to improve automated yield measurement in orchards and help the construction of smart orchards,this paper constructs an automated yield measurement model for apples based on current deep learning technology and popularly used smartphones to accurately and effectively calculate the yield of apple fruits.In the study,three different YOLO detection models were compared and analysed,and finally YOLOv7-tiny was selected as the detector,combined with Deep Sort multi-objective tracking algorithm to implement effective tracking of apples in the video stream,proposing a yield measurement model based on YOLOv7-tiny+Deep Sort algorithm and designing a simple and practical APP.research work and findings are as follows:(1)Apple dataset construction.This study used an unmanned aerial vehicle(UAV)and a mobile phone to acquire image data in the apple production area of Zhaotong,Yunnan.A total of approximately 1600 images were acquired in two apple patches,and the image enhancement technique was used to expand the dataset to 4697 images.The 4697 images were then annotated with Labelimg image annotation software and made into a YOLO format dataset,which was divided into a training set,a test set and a validation set in the ratio of 6:2:2 for the training of the model.(2)Apple counting study based on target detection algorithm and multi-target tracking algorithm.The YOLOv7-tiny,YOLOv3 and YOLOv5 s target detection models were trained on the dataset for apple detection.Through the performance comparison of the models,YOLOv7-tiny was finally chosen as the detection model,and this detector achieved an accuracy of 92.5%.The apple counting model was improved by replacing the detector with YOLOv7-tiny and tested to develop an apple counting model based on the YOLOv7-tiny target detection algorithm + Deep Sort target tracking algorithm,which achieved an average accuracy of 92.6% for apple counting.(3)Study of apple quality.A total of two quality calculation schemes were investigated:a calculation based on the ratio of images to actual apples,and a calculation based on the average quality of apples.A total of 506 sets of experimental data were tested in the scenario to calculate the ratio of the image to the actual apples in terms of feet of storage.Scheme 2tested a total of 1500 apple masses to calculate the average mass of Zhaotong Red Fuji apples.Finally,five sets of video control experiments were tested separately,and through analysis and comparison,the average accuracy of Scheme 1 was 83.32% and that of Scheme 2 was89.54%.The calculation scheme based on the average quality is more stable and has more advantages in terms of calculation speed.