Positioning Algorithm Of Machine Vision For Stacking Cartons

Intelligent mobile robot for loading and unloading replaces human to realize cargo loading and unloading automation,and one of its core technologies is the segmentation and positioning algorithm for stacking cartons,but few algorithms can meet the requirements of high recall rate in industrial scenarios.In this thesis,segmentation and positioning algorithm for stacking carton based on machine vision is studied to meet the demand of the loading and unloading robot.The main research contents are as follows:The technical route based on the object detection technology with deep learning is determined according to the overall needs and the work scenario of the loading and unloading robot.A carton dataset is constructed by collecting the images of stacking cartons in loading and unloading scene and warehouse environment to solve the lack of open source dataset in related scenario.A new labeling method which introduces carton attributes and gives fine-grained classification to the object is proposed.The characteristics of the dataset are analyzed quantitatively to support the model improvement.A new method of carton detection based on the prior information of the cargo and neural network is proposed which uses the position of the logo on the carton to assist to locate the carton.RetinaNet is chosen as the baseline after studying the performance of some classical object detection models on our carton dataset,which the segmentation head is added on to make the model utilize the mask annotations.Meanwhile,a logo segmentation model is built on basis of U-Net which uses one additional encoder network to extract the feature of logo template and utilizes multi-scale feature concatenation to realize the feature fusion between template and raw image,and dice loss is used to train the model.The pixel accuracy of the model on the test set reaches 99.65%,and its performance satisfies requirement for assisting carton detection.The RetinaNet’s deficiency on carton detection is improved and the integration deployment of the model on our robot based on ROS is completed.The IOU prediction subnetwork is introduced to directly realize the prediction of IOU for solving inaccurate quality prediction problem to the predicted bounding box.The anchor guiding module is designed to provide high-quality anchor for solving the mismatch between the shape size of the anchor window of the original model and the distribution of our carton dataset,so as to improve the accuracy of the predicted box position.The non-maximum suppression algorithm with linear weighting is used to alleviate the mutual suppression among different object boxes in densely packed scene,and min-IOU is introduced to replace the original IOU in the process of non-maximum suppression to solve the non-suppression of the false detection on small objects.The mAP of the improved model can reach 92.1%,which the recall rate on the test set can reach 99.7% when the mistake rate is less than 1% and can meet the requirements of industrial application.In addition to the positioning of stacking cartons,the scheme can also be used to segment and locate common cartons in other warehouse logistics scenarios,which has a broad application prospect.