Research Of Cotton Seedling Counting Method Based On Deep Learning And Multi-Object Tracking

In recent years,deep learning technology has developed rapidly and has broad application prospects in the field of agriculture.Cotton is an important cash crop and national strategic material in China.How to use deep learning technology to help farmers better carry out agronomic practice in the process of cotton production is an important research topic which has great research value and economic value.Seedling counting is an important work in agricultural production.The planting density calculated by the number of seedlings is closely related to the yield and quality of crops.The uniformity of seedling distribution determined by the number of seedlings can provide a basis for improving sowing equipment and technology.Therefore,it is necessary to carry out seedling counting in cotton seedling stage.However,the traditional manual statistical methods are usually time-consuming and inaccurate,and the seedling number regression model established based on the research of traditional image processing technology usually does not have universality.The research goal of this work is to develop a high-precision cotton seedling counting method using deep learning and multi-object tracking technology.The main research contents are as follows:(1)The CottonSeedling dataset and TAMU2015-ID dataset are constructed by collecting data in the field and manually labeling images,and the improvement of original open-source dataset respectively.(2)The anchor-based object detection method is studied.Based on Faster R-CNN and YOLOv4,two-stage cotton seedling object detection model Faster R-CNN-Cotton Seedling and one-stage cotton seedling object detection model YOLOv4-CottonSeedling are built respectively.The experiment results show that in the strict IoU(Intersection-over-Union)threshold condition,the performance of YOLOv4-CottonSeedling model is obviously better than that of Faster R-CNN-CottonSeedling model.At the same time,there is a problem that small-size targets are easy to be missed in Faster R-CNN-CottonSeedling model,which is well solved by YOLOv4-CottonSeedling model.(3)The anchor-free object detection method is studied.Based on the anchor-free object detection method CenterNet,the cotton seedling object detection model CenterNet-Cotton Seedling is built.The influence of backbone,transfer learning,learning rate attenuation strategy and attention mechanism on the performance of the model are researched respectively.According to the comparative experiment results,the improved model CenterNet-CottonSeedlingimproved is built and trained.The experiment results show that the performance of the model is significantly better than Faster R-CNN-Cotton Seedling model and YOLOv4-CottonSeedling model,which can satisfy the needs of detecting cotton seedling and distinguishing weed in video frames.The model is used as the object detection model in subsequent counting tasks.(4)The feasibility of counting cotton seedlings using multi-object tracking is studied.Kalman filtering and IoU metric based cotton seedling counting method SeedlingCountingbasic,the motion model and appearance model combined cotton seedling counting method SeedlingCounting-combined,trust mechanism and priority mechanism fused cotton seedling counting method SeedlingCounting-improved are developed separately.The SeedlingCountingimproved method has achieved r2(coefficient of determination)of 0.90,RMSE(Root Mean Square Error)of 0.6,and MAE(Mean Absolute Error)of 0.6 on CottonSeedling dataset,while achieved r2 of 0.97,RMSE of 1.33,and MAE of 1.32 on TAMU2015-ID dataset,which well meets the demand for accurate counting of cotton seedlings in the video.