Commodity Recognition System Based On Target Detection And Tracking

The emergence and rapid development of new retail industry has produced many new forms of unmanned entities.As a small and convenient entity,vision-based intelligent vending machines are gradually replacing traditionalvending machines,with low container damage rate for few mechanical devices,flexibility,convenice and no waiting time.And it has good development potential.In intelligent vending machine applications,the changing scenes lead to large changes in ambient light,the door movement during the opening phase makes the background unstable,there are many unrelated motion areas in the video image,and the goods are both taken and returned during the transaction.These characteristics lead to a low commodity recognition rate in this scenario.At present,there is no perfect technical solution.Therefore,a stable and reliable commodity recognition solution needs to be designed according to the actual scenario requirements.This article designs a commodity recognition system based on target detection and tracking,which is used to describe the commodity transaction process of intelligent vending machines and evaluate the transaction results.There are a large number of invalid areas both inside and outside the container in the images of the actual container commodity transaction video.By selecting a Region of Interest(ROI),motion interference in the invalid area is eliminated,and the system calculation is also reduced.After the image is collected,the intelligent vending machine will go through a period of time during which the door is opened.It is often very short to purchase goods after the door is opened.It is difficult to establish a stable and quickly updated background model.This paper analyzes the principles of three moving target detection algorithms,and proposes an optimized three-frame difference method based on shape features to detect moving targets.The shape features of moving targets are used to filter out irrelevant separated movements.The holes generated by frame differences are used to eliminate the motion of the arm connected to the commodity.Considering that the commodity area will also decrease at the same time,a window optimization method based on the skin color is proposed to convert the independent commodity detection problem into a problem with skin color-dependent characteristics,obtaining more accurate product information and product positioning,and track it.And a matching function is designed to verify the accuracy of commodity tracking.This paper also proposes a behavior judgment method based on trajectory smoothing,which estimates the motion states and combination them to divide behavior patterns,avoiding repeated recognition and misrecognition of commodities.Finally,based on the self-built commodity data set,the product image is classified using the Inception V3 classification network,and the commodity recognition result is given in combination with the behavior pattern.This paper tests the commodity recognition system through commercial videos.In the video test of taking a single commodity and that of taking multiple commodities,the recognition accuracy reached 74.5% and 70.7%,respectively,which has good and stable performance.Among them,the accuracy of behavior judgment of detected target is 80.8%and 73.2%,respectively,which proves that the detection and behavior judgment methods proposed in this paper are robust to the complex application scenarios of intelligent vending machines.