Design And Implementation Of Parcel Singulation System Based On Machine Vision

With the rapid development of China's e-commerce industry,online shopping has incereasingly become the main way of shopping.Facing hundreds of millions of express orders every day,how to improve the efficiency of parcel sorting and distribution has become an urgent problem for all express logistics companies.This paper focuses on the problem of parcel singulation in the fully automated express parcel sorting production line,and we designed and implemented a parcel singulation system based on vision.We designed and implemented two express parcel segmentation algorithms,which effectively solves the problem that express parcels are difficult to segment.We optimized the inference scheme of express parcel segmentation algorithm,so that the system was successfully deployed in the embedded development board.The main work of this paper is three-fold:1.Designing and implementing a parcel singulation system based on vision.We designed the overall scheme of the system in detail,and we introduced the various components of the system and its functions,and designed four software modules,including image acquisition module,image preprocessing module,express parcel segmentation module and singulation control module,and also designed a simple and easy-to-use front-end user interface.The system has been used in the fully automated express parcel sorting production line of a sorting center of an express company.The separation area is composed of 4 × 8 small conveyor belts with length and width of500 mm and 150 mm respectively.Under the condition of controlling the parcel spacing of 1000 mm,it can separate about 4500 parcels of different shapes and sizes per hour,which is at the advanced level of similar products in the industry.2.Designing and implementing the depth image express parcel segmentation algorithm based on connected region markers and the RGB image express parcel segmentation algorithm based on YOLACT.From the three aspects of the algorithm's inference time,the average accuracy of the detection box and the average accuracy of the segmentation mask,we compare these two segmentation algorithms on the selfmade express parcel data set.We also analyzed the applicable scenarios of the two segmentation algorithms,which provides guidance and basis for the application of the actual system.3.Designing and implementing a solution to optimize the network model structure of express parcel segmentation and accelerate its inference speed.In this paper,we use the Tensor RT framework to optimize the network model structure of the express parcel segmentation,and use low-precision data to calibrate its network model weight parameters to accelerate the network inference process.On the NVIDIA AGX-Xavier embedded development board,we use the self-made express parcel data set to evaluate the network model before and after optimization.For the unoptimized network model,the average precision of segmentation mask is 88.80 m AP and the inference speed is8.93 FPS.For the optimized network model,the average precision of segmentation mask is 88.27 m AP and the inference speed is 31.57 FPS.Experimental results show that this solution can greatly improve the inference speed of the network model while losing low precision.The parcel singulation system based on vision designed and implemented in this paper has been successfully applied in a sorting center of an express company.The system improves the automation and intelligence of express parcel sorting,and provides an effective guarantee for improving the delivery efficiency of express parcels.