Carton Dataset Construction And Research Of Its Vision Detection Algorithm

China's logistics technology and logistics industry have developed rapidly in recent years,but inefficiency and high cost still remain the prominent problems.Applying intelli-gent logistics technology is one of the most important ways to solve the problems.Carton detection algorithm is an important and fundamental intelligent logistics technology and can be widely applied to many logistics tasks such as unloading,destacking and accessing of cartons.However,the research of carton detection algorithm is faced with the challenges of the lack of carton dataset and the high localization accuracy demand in logistics tasks.Therefore,the construction of carton dataset,loss function,model structure and learning strategy are studied in this thesis.The main contributions and conclusions are as follows.(1)To solve the problem of the lack of carton dataset,we design the principle of data collection,cleaning,labeling and construct the first publicly available carton detection dataset SCD.SCD has the characteristics of large scale,rich diversity,high-quality labeling infor-mation and dense stacking of goods.This dataset provides an important evaluation bench-mark for the research of carton detection algorithm and can greatly promote the research and application of carton detection.(2)To solve the problem of the low correlation between classification score and local-ization accuracy caused by independent losses and the domination of outliers' gradient in the localization loss,we propose Io U-balanced classification loss and localization loss.The former assigns adaptive weight for the classification loss of each positive example based on Io U and drive the models to learn high classification score for examples with high lo-calization accuracy and learn low classification score for examples with low localization accuracy.Thus,the correlation between classification score and localization accuracy can be enhanced.The latter assigns adaptive weight for the localization loss of each positive ex-ample based on Io U,which alleviates the problem of the domination of outliers' gradient in the localization loss.The experimental results on the carton detection dataset and common object detection datasets show that the proposed method can greatly improve the localization accuracy of carton detection models without hurting the inference efficiency and has good generalization ability on the common object detection task.(3)To further solve the problem of the low correlation between classification score and localization accuracy caused by independent classification and localization branch,we re-design the architecture of detection models and propose Io U-aware object detection models.This method attaches an Io U prediction layer parallel to the localization layer to predict lo-calization accuracy for each detection.During test,the detection confidence is computed by multiplying the classification score by the predicted Io U.Thus,the detection confidence is more correlated with the localization accuracy and used as the input of NMS and AP com-putation to improve models' performance.The experimental results on the carton detec-tion dataset and common object detection datasets show that this method can further greatly improve the localization accuracy of carton detection models and has good generalization ability on the common object detection task.(4)To solve the problem that the standard Io U pays insufficient attention to the close-ness between the center of anchor box and ground truth box during training example sam-pling,we design a better metric Gaussian guided Io U(GGIo U)and propose GGIo U-based balanced learning method.Compared with Io U,GGIo U focuses more attention on the close-ness between the center of anchor box and ground truth box.GGIo U-based balanced learn-ing method can assign more training examples to the slender objects and drive the learning process to focus more attention on the features whose receptive fields align better with ob-jects.The experimental results on the carton detection dataset and common object detection datasets show that the method can largely improve the localization accuracy of carton detec-tion models without hurting the inference efficiency and has good generalization ability on the common object detection task.(5)With the carton detection dataset SCD and the above proposed methods,we pro-pose the carton detection models with high localization accuracy and then design the vision detection system of cartons for the logistics.The proposed vision system is applied to the in-telligent loading and unloading mobile robot and the effectiveness of the proposed algorithm is verified.