Research On Multi-Object Tracking Algorithm For Soybean Plant Test

Soybean is an important source of high-quality protein and edible oil for human.In recent years,the demand for soybean in China has increased rapidly with the increase of population.However,most of the soybean supply depends on imports,which seriously threatens the national food security.Therefore,it is urgent to achieve high quality and yield of domestic soybean.One way of soybean plant testing is to identify and analysis of soybean plant phenotypic traits,which is of great significance to improve the yield of soybean.At present,the detection of soybean plants mainly relies on traditional artificial method,and the phenotypic character recognition mainly focuses on pod species detection and bifurcation number research.There is no effective method to predict seed number per plant.In this thesis,we propose a multi-object tracking(MOT)frame based on the research object of the Northeast soybean plants after harvest.Soybean plant rotation video are collected as input data by soybean single plant phenotyping instrument.The introduced method realizes the automatic prediction of soybean pod per plant and soybean seed quantity.It also provides a reliable basis for evaluating the yield of different varieties of soybeanThe main work of this thesis is as follows(1)A soybean single plant phenotype measuring instrument is designed and manufactured by combining industrial camera,adjustable plant clamping structure,mechanical arm and stilling lamp.More than 700 double plant rotation videos are collected from different varieties of Northeast soybean harvested in 2020.The object detection(OD)and tracking data set is realized through image clipping algorithm,image format conversion algorithm and Labellmg manual annotation.(2)Based on the YOLOv4 model,the Improved-YOLOv4 model is proposed by introducing the k-means clustering algorithm and the improved attention module,which effectively predicts the number of pods per soybean plant.After a large number of experiments,it is testified that the mean average precision of the detection model in this thesis is 5.67%higher than that of YOLOv4,84.37%in the extended data set,and 99.1%in the pendulum data set.The detection accuracy is greatly improved,and the robustness and generalization ability of the model are verified(3)Based on the DeepSORT model,a multi-category pod tracking model called MBI-DeepSORT is proposed.It combines the improved YOLOv4 model and trains the depth of the apparent characteristics of bean re-identification(B-ReID)based on the idea of transfer learning.A multi-category tracking and counting module is added and the function of soybean pod tracking per plant and soybean seed quantity statistics is realized.The introduced frame provides a new way for artificial intelligence breeding,and completes one of the core tasks of soybean seed test per plant.In this thesis,a new framework of soybean pod detection,tracking and counting is designed based on object detection and multi-object tracking technology.The preliminary experiment of soybean seed test is completed,which effectively alleviates the time-consuming,laborious and large error of traditional seed test.It provides a reliable data basis for breeding experts to evaluate the yield of different varieties of soybean,and provides an effective prior method for automatic seed test of soybean plants.