Research On Intelligent Parcel Sorting System In Complex Scenarios Based On Deep Learning

At present,large-scale logistics sorting centers mostly use the sorting method of flow operations represented by cross-belt sorters,but most of the links such as express information screening at the sorting entrance and express sorting and transportation at the exit are manually performed,resulting in The problem of low sorting efficiency and high error rate greatly limits the efficiency of logistics transportation.In order to solve the above problems,this paper designs and builds an intelligent express sorting system,which effectively improves the sorting efficiency.The specific content and research results are as follows:First,to address the problem of low accuracy of express delivery detection and recognition in complex scenes with disorderly stacking and mutual occlusion,a new target detection and recognition algorithm based on channel information fusion is proposed.The algorithm combines the multi-layer shallow feature map and the final feature map with interlace and column downsampling method for channel fusion,so that the feature map has rich semantic information and spatial information at the same time,extracting more subtle features.Through experimental verification,the detection accuracy of the new target detection and recognition algorithm is 98.63%,which is about 2%higher than the original target detection algorithm.Secondly,to solve the problem of time-consuming determination of target picking poses in complex scenarios,a cascaded convolution-based optimal picking position estimation algorithm based on key points is proposed.The algorithm determines the optimal picking pose of the object in real time through the optimization of the cascade network,using the key point algorithm and the correlation correction loss function,and improves the robot picking efficiency.Through the experimental verification on the CPU,the algorithm can be used to determine the position and posture of the target in real time,which is about 26FPS.Thirdly,in view of the low sorting efficiency of the robot sorting system,a model pruning algorithm based on dynamic scale variables is proposed.The algorithm firstly combines model weights and dynamic scale variables for model sparse training and automatically recognizes the importance of the channel,and then uses a pruning strategy to generate a lightweight model with comparable accuracy.Through experimental verification,the lightweight detection model reduces memory consumption by 33.9%and improves the detection speed by 18.63%at the expense of 2.17%accuracy.Finally,the hand-eye calibration and the design of the robot control instructions are carried out,and the relationship between the coordinate system of the image acquisition module and the robot is uniformly expressed.Secondly,in view of the low accuracy of express delivery detection and recognition in complex scenarios,a new target detection and recognition algorithm based on channel information fusion is proposed.The algorithm combines the multi-layer shallow feature map and the final feature map with interlace and column downsampling method to channel fusion,so that the feature map has rich semantic information and spatial information at the same time,extracting more subtle features.Through experimental verification,the new target detection and recognition algorithm improves the accuracy of detection and recognition.Thirdly,to solve the problem of time-consuming determination of target picking pose in complex scenes,a cascaded convolutional optimal picking position estimation algorithm based on key points is proposed.This algorithm determines the optimal picking pose of the object in real time through the optimization of the cascade network,using the key point algorithm and the correlation correction loss function,and improves the robot picking efficiency.Through experimental verification,the algorithm can be used to determine the position and posture of the target in real time.Finally,based on the fact that deep learning consumes a lot of computing resources and there are complex scenes in which the sorting process is piled up in disorder and mutual obstruction,an intelligent express sorting system verification platform is built to achieve express sorting in complex scenarios.The intelligent express sorting system is divided into a control layer,an application layer and a cloud layer from a logical level,so that each module of the system has the characteristics of strong cohesion and loose coupling,which is convenient for the subsequent development and promotion of each module.Combined with the application scenarios of the target detection algorithm,each module of the control layer,application layer and cloud layer was selected.