Kiwifruit Yield Estimation Based On YOLOv7 And Deepsort Algorithm

As one of the key links of agricultural production management,fruit yield estimation not only helps to rationally allocate labor,but also helps agricultural producers effectively use labor resources such as storage space and picking equipment.With the acceleration of modernization,there is a large gap between labor and agricultural demand,so it is important to realize smart agricultural production management.In this paper,aiming at the characteristics of dwarf dense planting and wide fruit distribution of kiwifruit cultivated in greenhouses,this paper used orchard crawler trolleys to shoot kiwifruit orchards,and combined artificial labeling software to establish a dataset of kiwifruit detection and tracking,and proposed a yield estimation scheme based on YOLOv7 and Deepsort algorithms.YOLOv7+Soft-NMS detection:K-means clustering was used to obtain 9 types of prior boxes,and the dataset was enriched by data augmentation,and the preprocessed data was detected by YOLOv7 model plus improved Soft-NMS to detect kiwifruit in the image and compared with different object detection algorithms.The experimental results showed that the improved YOLOv7 model performs well on the kiwifruit detection dataset,with an F1score of 92.60%and a large video processing speed of 25.84FPS.Deepsort target tracking:The trained YOLOv7 network was used as the detector of the Deepsort network to track the shooting video,respectively,the VGG16 network was used to extract the appearance features of the target,the Kalman filter extracts the motion features of the target,and the Hungarian algorithm completes the target matching of the frames before and after the video,and the Deepsort algorithm was jointly constructed,and the average tracking accuracy was 89.87%and the average tracking precision was 82.34%.ID yield counting:The collision line estimation based on YOLOv7 detection algorithm and the ID counting based on YOLOv7 and Deepsort tracking algorithms were used to estimate the yield of 7 videos in simple,normal,and difficult modes.Compared with the obtained ID counting strategy,the accuracy of the highest estimation can reach 97.49%.In order to further improved the yield estimation accuracy of kiwifruit dense areas,the manual statistical results were fitted with the ID counting results by linear regression,and the correlation coefficient R~2 was 0.9887,and the average accuracy of ID counting in difficult mode was increased by 11.53%.