Research And System Implementation Of Fast Target Object Recognition In Open Environment

Object recognition is an important technology in the data sensor layer of the Internet of Things(IoT).Image recognition technology of obj ects obtains information from the perspective of human vision,eliminates part of the "blind area" of traditional object recognition technology,and obtains more useful information.It can also be used as a supplement to traditional object recognition technology such as RFID to enhance the scene sense ability of IoT technology.In the field of logistics,civil camera equipment is widely used,and the non-compatibility of image recognition also exerts its special advantages in some situations.In recent years,with the rise of deep learning,the accuracy,robustness,recognition efficiency and application scope of general object recognition methods have been greatly improved,but there are still some difficulties and problems in the open environment at this stage.In order to take control of the real-time status of goods in retail stores,make full use of abundant scene data,improve the efficiency of commodity management and customer experience,this paper aims at the task of automatic commodity recognition in open environment,trying to improve the performance of commodity recognition in open environment.The main work finished is as follows:(1)Firstly,from the perspective of scenarios and data,the business scenarios of commodity identification in open environment are analyzed.40 kinds of commodities with different appearance characteristics are selected as target objects for recognition.The training data acquisition schemes for different commodities are designed,and a large number of test data close to the actual open scene are collected for this experiment.An open environment commodity recognition data set containing 1461 images was constructed.(2)In order to ensure the real-time requirement and low power limitation of this project,the resolution impact of the current mainstream target detection framework,as well as the accuracy and time performance indicators are analyzed,and the identification model based on YOLO framework is determined.The feature extraction ability of candidate networks is investigated through classification experiments,and DarkNet-53 is chosen as the lightweight backbone network for the task of this paper,which improves the detection efficiency while guaranteeing the accuracy of the model.(3)In order to alleviate the cross-distribution problem of commodity recognition in open environment,a data augmentation strategy combining multiple geometric transformation operations and Mixup is designed.By introducing the strategy of fine-grained feature recognition,the model can obtain more meaningful semantic information from the fine-grained information of the up-sampled features and the early feature mapping,so as to improve the detection ability of the model for dense small objects in open environment.Experiments show that the proposed strategy improves the model by 3.9%and 3.4%on the two test sets respectively.After adjusting the data,the mAP of the commodity identification model is also significantly increased from 80.68%to 89.42%.(4)Finally,on the basis of the proposed object recognition model in open environment,the system is implemented.The system finally achieves the functions of video stream acquisition,commodity recognition,visualization of results and archiving.It can be used as the basic module of commodity intelligent recognition analysis technology and bring new opportunities for intelligent retail.