Investigation And Application Of Human-Object Interaction Detection Algorithm

With the development of computer vision in basic tasks,such as target recognition,the research direction of computer vision is gradually turning to higher level image semantics and practical engineering applications.Extracting the high-dimensional semantic of an image requires not only the attributes of the objects in the image,but also the interaction between the objects.In most scenes and applications,people are the interaction center of image semantics,therefore,determining the interaction between people and other objects in the image is the most important key to understanding image semantics in engineering applications.Aiming at the human-object interaction detection,this paper studies the shopping behavior detection in unmanned convenience store,and further studies and improves the method of human-object interaction in general scenes.The work of this paper mainly includes the following two parts:(1)Implementation of unmanned settlement system based on camera and grating sensor.In this paper,an unmanned settlement system based on camera and infrared grating sensor is designed and implemented to identify the shopping behavior of unmanned convenience stores.In this paper,the intelligent shelves are linked with grating sensors by vision techniques such as human posture estimation,face recognition,target classification and so on,and the unmanned settlement system is realized.Experiments show that the unattended settlement system implemented in this paper has the advantages of lower hardware cost,less computation,flexible deployment and certain anti-theft capability under the condition that the precision is guaranteed.In this paper,a semi-supervised network training method is proposed to solve the problems such as large quantity of goods,fast iteration and high cost of manual labeling in unmanned convenience stores.Through the iterative training of the data,the recognition accuracy of the network can be improved without adding the labeled data,and the workload of the data tagging can be greatly reduced.Experiments show that the proposed method can effectively reduce the labeled data required for network training in both open datasets and handheld commodity data.(2)The investigation of the detection of interaction between human and object in combination with human pose estimation.The key point of human body posture is the important information to identify the movement or behavior of the human body.If the key point information can be added to the detection network,the effect of detecting the interaction between human and object can be improved by the network.Therefore,on the basis of HO-RCNN network framework,this paper proposes two methods to incorporate the key information of human posture.Key marks are added to the image data,and the spatial position information of the key points and the object detection box are added to the network to merge the key information into the key information.The experimental results show that this method can effectively improve the recognition accuracy of the network.