| With the development of high and new technology with computer as the core,the cost of robot has been greatly reduced.It is the future development trend to harvest fruit with picking robot.The main technology of the autonomous robot has been the rapid growth of target recognition using deep learning in recent years.In view of the fact that picking robot needs to judge the maturity of fruit in orchard,This paper proposes an algorithm for identification of fruit maturity and classification based on a convolutionary neural network.and verifies the applicability of the algorithm by experiments.At the same time,in order to solve the problems of low detection accuracy,slow speed of picking robot in the case of multi-target,small target and overlapping fruit in natural environment,this paper proposes a target detection and capture planning method based on improved deep learning algorithm,so that the picking robot can detect and grasp the fruit target more efficiently and accurately.Firstly,research on fruit maturity recognition and classification.A variety of fruit images were collected and data sets were constructed,and the image data were expanded by data enhancement technology.In this paper,an improved convolutional neural network model is proposed to recognize and classify fruit maturity,at the same time,the fruit maturity distinguition experiment was carried out.The whole experimental steps include: preparing data set,designing convolutional neural network,model training and model testing.In the test data set,The overall accuracy of identification for five fruit forms is 87.6 percent,and the highest recognition accuracy is 94%.Based on the standard,the precision is greater and the level of convergence that can be used for the detection and classification of selecting robots is higher.Secondly,researching the fruit image target detection.An improved SSD target detection model is proposed.The structure and loss function of SSD are improved.In order to check the effect of the improved model on identification,in this paper,four SSD models based on vgg16+without adding centerloss,based on improved resnet101+without adding centerloss,based on vgg16+with adding centerloss and based on improved resnet101+ with adding centerloss were compared and tested,The test results show that the SSD model based on the improved resnet101 wih adding centerloss is best,the loss is stable in the range of[0.25,0.1],which can be used for picking robot target detection.Finally,In order to verify the effectiveness of the target detection proposed in this paper,an experimental platform is built for objected detecting and grasping experiments.The objected detecting experiment shows that the improved target detection algorithm is not affected by the size of the target in the image,whether it is a short-range large target or a long-distance small target,the m AP can reach92.74% on the test set,and the single tested image only takes 0.72 s,it can effectively detect the fruit in the natural environment.The grab experiment shows that the average absolute error is 3.83 mm.The success rate of anno manipulator selected in this paper is 80%,which proves the practicability of the selected manipulator.In this paper,based on the practicability of picking robot target detection and grabing effect,an effective solution for intelligent fruit picking is proposed. |
