Research Of Fruit Picking Localization Algorithm Based On Deep Learning

In recent years,C hina is vigorously developing agriculture and forestry,and the output of agricultural and forestry crops is gradually increased,which has led to a sharp increase in the workload of fruit picking.Due to the low efficiency of manual operations,it is no longer possible to meet the actual demand during the harvest season.Therefore,the method of using robots for operation arises at the right moment.Agricultural robots have become one of the most promising.Many agricutural jobs are repetitive,such as so wing,weeding,and pruning.Agricultural robots solve repetitive,time-consuming tasks for farmers,freeing their hands.The robot has the characteristics of long service life,good repeatability and can work in bad weather.The development of robotic picking involves many challenging tasks,such as fruit selection and positioning.Therefore,the development of accurate fruit recognit ion and positioning system is an important step towards realizing automatic picking robot.At present,the algorithm applied to identify and locate fruits in fruit garden has some problems,such as undetected small fruits and undetected blocked fruits,result ing in unsatisfactory results.This article focuses on the above-mentioned key issues,focusing on the fruit picking positioning research.An improved algorithm for identifying fruit picking targets is proposed,which mainly solves the problem that small fruits and covered fruits cannot be identified.Based on the convolut ional neural network,this algorithm adopts the method of improving feature extraction mapping and increasing feature pyramid network according to the environmental characteristics of fruits and orchards,so as to obtain more detailed features and improve the detection rate of smaller fruits.In addition,the removal strategy of detection box was improved to avoid the problem of missing detection due to overlapping fruits.The fruit image acquis ition results are two-dimensional plane images,and the location of the target fruit located by the detection algorith m is also the location of the fruit in the picture.However,the actual three-dimensional coordinates of the fruit should be known before picking.In this paper,the strategy of establishing mult iple coordinate systems is proposed to solve the problem that the position of fruit in real space cannot be determined by two-dimensional plane.The experimental results show that the improved algorithm improves the accuracy of fruit target recognit ion,especially in the detection of small target fruit and overlapping fruit.The overall fruit recognit ion accuracy increased by 3.48 percentage points.