| China is a large country of apple production and consumption,and apple picking has been bothering farmers.Apple picking is seasonal and labor-intensive.At this stage,apple picking in China mainly depends on manual work.Intelligent picking can reduce costs,improve picking efficiency and increase economic income.It is an inevitable trend of apple picking.The accuracy and recognition speed of the image recognition system of apple picking robot directly affect the picking efficiency.In order to improve the accuracy of Apple recognition and positioning in the mature period of apple orchard,this paper selects convolution neural network as the basis,refers to the relevant knowledge and research results,and obtains the three-dimensional coordinates of apple through binocular stereo depth camera,so as to improve the accuracy and robustness of Apple target recognition and positioning of apple picking robot,and facilitate the follow-up apple picking robot to complete the picking work efficiently.The results of this study are as follows:（1）Research on apple image preprocessing.Due to the inconsistency of image size in the acquired apple image data set and the difference of image brightness and darkness,the experimental Apple data set is enhanced by using the methods of size normalization,random rotation,color jitter,gamma transform and motion blur to increase the training samples and improve the robustness and generalization ability of the model.The apple image was calibrated by labelimg software,and the apple data set was established.（2）Research on Apple target recognition algorithm.Firstly,aiming at the problem of insufficient detection accuracy,detection speed and generalization ability of fast r-cnn and other networks in traditional image processing target detection methods and deep learning algorithms,an improved yolov5 detection model is proposed for Apple target recognition.On the Yolo algorithm,the backbone network,network lightweight and bidirectional feature pyramid network are improved.By adding attention mechanism to the final model,the model can obtain the importance of different channel features through automatic learning;The repovgg neural network operator is used to improve the Yolo backbone network,which improves the detection speed on the premise of ensuring the accuracy of the algorithm,and meets the real-time requirements of the picking robot for the target fruit detection;Secondly,the bifpn bi-directional feature pyramid network is combined with the yolov5 target detection framework to fuse more features without increasing too much operation cost.In order to improve the accuracy of small target detection,we add a new feature layer to achieve a higher level of feature fusion and improve the accuracy and speed of Apple target recognition.The accuracy of the improved Yolo algorithm for Apple target recognition is 0.973,and the recognition speed is 0.032 frames / s.The method proposed in this paper has significantly improved the detection accuracy and speed,and the apple recognition effect is better.（3）By calibrating the binocular cameras,the internal parameter matrix of the left and right cameras is obtained.Based on the principle of polar constraint,the pixels of the left and right images are matched.The obtained depth data are processed by bubble sorting and median filtering.The two-dimensional coordinates of the apple are converted into three-dimensional coordinates by parallax method.The three-dimensional information of the apple fruit is obtained by similar triangle operation.The analysis of the experimental results shows that the average depth distance error between the distance measured by the binocular camera and the actual distance is 15.9mm,and the measurement error is small,which meets the general picking requirements. |
