Research And System Design Of Commodity Identification Algorithm In Intelligent Container Scene

In recent years,with the rapid rise of artificial intelligence,unmanned retail has gradually developed,among which intelligent container has become a direction for e-commerce companies to pay close attention to and all kinds of capital to pursue.With the development of deep learning in the field of machine vision,intelligent container has been constantly improved,transforming from a non-visual traditional solution to a vision solution based on deep neural network.In this paper,based on the visual scheme of intelligent container,the identification of goods is studied.The main work is as follows:(1)To solve the problem of missing standard data set of intelligent container,static scene data set FSD20 and dynamic scene data set FDD20 are respectively established according to the characteristics of the scene.Among them,the static data set FSD20 contains 88 categories of commodities and more than 8,300 pictures.Dynamic scene data set FDD20 has 15 video sequences of picking up goods,including 10 categories of goods.(2)For static scenes,this paper proposes a static scheme based on detection.Detection model in order to reduce the static scheme in the embedded devices on the amount of calculation,improve speed of reasoning models,this paper USES bottleneck residual block of Center Net network residual block was improved,and put forward Center Net-D(Center Net-based on Depthwise Separable Convolution)model structure.On the basis of centernet-D,this paper further constructs the Centernet-DD(Centernet-D based on Dilated Convolution)model,and increases the network sensation field without increasing the number of parameters,thus improving the detection accuracy of large targets.The experiment on static state scene data FSD20 shows that centernet-DD algorithm reduces the number of parameters from 15.7M to 3.5M,and the number of floating-point operations from 6.14 G to 0.66 G,and the accuracy rate is improved to some extent.(3)In view of the shortcomings of the static scheme,such as large number of cameras,low space utilization rate and limited positioning,an exploratory dynamic scheme based on multitarget tracking was proposed and implemented.A lightweight web Bottle Net6 is designed to rapidly extract depth appearance characteristics of products in multi-target tracking.In dynamic scheme,based on the improved Center Net-DD algorithm added movement features and the depth of the Bottle Net6 extraction,and to identify and based on Kalman filter in combination with goods tracking,experiment comparison found that adding motion features and depth can effectively reduce the residual and deterrent to commodities,improve the stability of intelligent container goods identification.To sum up,this paper firstly proposes and realizes the static scheme based on detection for the commodity identification of intelligent container,and proposes and realizes the dynamic scheme based on multi-objective tracking for the deficiency of the static scheme,which achieves better practical effect and has strong practical application value.