Research Of Human-object Interaction Detection Based On Deep Learning

The development of the Internet has brought a lot of complicated image data,which has led to a large number of sub-topics arising from computer vision.With the development of deep learning in the field of computer vision,many topics have achieved good results,such as image recognition,face detection and recognition,object detection,etc.,but some complex topics,like human-object interaction(HOI)detection,have not yet made a breakthrough.The information age is most concerned with humancentered image understanding,and human-object interaction detection has a wide range of applications.This paper focuses on the existing deficiencies of human-object interaction detection.The main research work is as follows:(1)In view of the fact that the existing main methods ignore or fail to make effective use of the local details of human and objects in the image,a Human-Object Interaction Recognition Model Based on Attention Mechanism(HOIR-AM)is proposed.The method firstly uses the attention mechanism to extract the visual features,so that the visual features not only include the features of the human and object instance,but also the local features of interest of the human and the object,and provide more fine-grained effective information for distinguishing various interactions.Then the method extracts the location features from the relative positions of human and objects,and finally uses the fusion of the visual features and the location features to identify the interaction between the human and the object.With the object detection algorithm,our results show that the proposed method achieves 46.62 mAP on V-COCO dataset.(2)For the problem that the matching efficiency between the human and the object instance which detected by object detector combination is low in the existing method and the method in(1),a Detection of Human-Object Interaction Based on HumanObject Interaction Matching Network(HOID-HOIMN)is proposed.The method firstly uses the interaction matching network to identify the interaction between human and objects,and then uses HOIR-AM method to complete the interaction recognition.With the object detection algorithm,our results show that the proposed method achieves 46.69 mAP on V-COCO dataset,which is higher than HOIR-AM,and the method recognition time is 12.7% less than HOIR-AM.(3)For the problems of cumbersome and time-consuming detection steps in(1)and(2),a System of Human-Object Interaction Based on HOID-HOIMN is designed and implemented.The system solves the problem of feature extraction by using feature sharing,and target detection,interaction matching and interaction recognition in a single framework are unified under one framework,realizes one-step HOI detection,and achieves a detection time of 0.2 s/f,which has good practicability.