Orchard Apple Detection Based On Computer Vision In Natural Environment

Apple picking is an important part in apple production.Traditionally,manual picking usually has problems such as waste of manpower and material resources,low efficiency and difficult to ensure the personal safety of workers.Therefore,people realize agricultural automation through fruit picking robots to achieve the purpose of safely and efficiently completing the fruit picking task.When the robot performs the fruit picking work,it must have good detection performance,so as to accurately locate the fruits and improve the work efficiency.In practical work,the natural environment is complex and changeable,such as lighting conditions,branches and leaves occlusions and other factors will affect the accuracy of apple detection.In order to achieve effective and stable orchard apple detection and improve the production efficiency,based on the existing object detection methods,this thesis carries out the research on orchard apple detection based on computer vision in the natural environment.The main work is as follows:1.Study on apple detection in orchard based on Conditional Boosted Random Ferns(CBRFs).In order to improve the robustness of traditional orchard apple detection methods under the influence of complex conditions such as illumination and scale transformation,this thesis takes Boosted Random Ferns as the basic algorithm and proposes a detection algorithm based on Conditional Boosted Random Ferns.Firstly,the HOG(Histogram of Oriented Gradients)feature of apple dataset image is extracted to construct the feature space.Secondly,in order to enhance the ability of features expression,binary features are randomly selected in different direction intervals,and the correlation between features and direction regions is introduced to generate binary features and construct random ferns.Then,the weak classification is constructed by random fern and the target distribution probability is calculated.Finally,the Real Ada Boost enhancement strategy is used to select the most discriminative weak classifier and construct a strong classifier for classification prediction.It is verified on the self-made apple dataset and the international public acfr dataset.The comparative experiments show that the algorithm can effectively improve the accuracy of apple detection in orchard in natural environment.2.Research on apple detection in orchard based on improved Faster R-CNN.Compared with traditional methods,the Deep Learning method does not need to design features manually,and the model has strong adaptive ability.This thesis proposes an improved apple detection method of Faster R-CNN,takes Faster R-CNN network as the basic model for improvement and optimization,constructs Res Net50 +FPN(Feature Pyramid Networks)backbone network,extracts features,uses multi-scale anchors generator to frame targets in the Region Proposal Network(RPN)module,and introduces Soft-NMS(Non-Maximum Suppression)algorithm to eliminate redundant region suggestion boxes.In order to reduce the classification and regression positioning error of the regional suggestion frame caused by the loss of spatial symmetry,the ROI Align(Region of Interest Align)method is used to scale the candidate regions generated by the RPN module to the corresponding scale feature map,and the classification regression prediction is carried out through the full connection layer to enhance the detection performance of the model.Experiments on self-made apple datasets and international public acfr data sets show that the improved Faster R-CNN apple detection algorithm has high detection and recall.Therefore,the detection algorithm has certain advantages in target detection tasks in the natural environment.Meanwhile,experiments show that the detection speed reaches 5fps,which can meet the needs of real-time detection in application.To sum up,the two algorithms proposed in this thesis enrich the basic theory of target detection in complex environment,and provide a reference for improving the efficiency of apple detection in orchard and the technical ability of related industries,which has a certain significance and a practical application value.