| Human-object behavior analysis technology is an important research direction in the field of computer vision of artificial intelligence.Deep learning is widely used in human-object behavior analysis because of its learning ability.However,there are still problems in the current human-object behavior analysis technology,such as insufficient utilization of interaction features between human-object pairs and unclear or missing human-object behavior objects in complex backgrounds.In view of these difficulties,this paper studies the key technologies of human-object behavior analysis based on deep learning.The specific contents are as follows:(1)Aiming at the problem of insufficient utilization of interactive feature information between human and object,this paper proposes two-stream Networks based on interactive feature generation algorithm.The dual-stream network extracts humans’ and objects’ feature information and human-object interaction feature information respectively.When extracting humans’ and objects’ feature information,considering the powerful feature expression ability of the ResNeXt,the network is used to extract the humans’ and objects’ features in the image;when extracting the interactive feature information between human and object,the feature reasoning ability of the graph convolutional neural network is used to propose interactive feature generation algorithm.The features of humans and objects are used as nodes,and the interaction between human and object is used as edges to construct a graph model.According to the interactive feature generation algorithm,the input graph model is updated by traversing the nodes,and new interactive features are generated during the updating process.Finally,the humans’ and objects’ feature information and the human-object interaction feature information are fused and sent to the classification network for behavior recognition,so as to fully utilize the humans’ and objects’ feature information in the image and the interaction feature information of the human-objects.(2)Aiming at the problem of unclear or missing human-object behavior objects in complex backgrounds,this paper conducts research on the constructed two-stream Networks,and proposes a human-object behavior analysis method based on interactive feature generation and key human-object enhancement.Considering the lack of human-object behavior objects in complex backgrounds,the OFR-CNN(Optimized Faster Region Convolutional Neural Network)is proposed to solve the problem of difficult detection of small objects and occluded objects.The network uses a multi-scale feature optimization algorithm to optimize the Faster R-CNN network.The shallow features in the ResNeXt feature extraction network are subjected to maximum pooling,the deep features are deconvolved,and the feature layers are fused to obtain the final image features.Considering the unclear problem of human-object behavior objects in complex backgrounds,this paper proposes a key human-object enhancement algorithm to solve the fuzzy problem of target human-object pairs in the analysis process.The key human-object enhancement algorithm is used to enhance the spatial position attention of the acquired human-object feature target frame,traverse the image target to find the target human-object pair to form a weight matrix,and multiply the obtained image feature point to obtain the feature information required for the final recognition.The human and object features obtained by the improved network and the human-object interaction features obtained by the interaction generation network are fused to obtain the final feature information,which is sent to the classification network to obtain the human-object behavior analysis results.In this paper,the proposed human-object behavior analysis method based on interaction feature generation and key human-object enhancement is experimentally verified on the HICO-DET dataset and the V-COCO dataset.The results show that compared with the existing human-object behavior analysis technology,the accuracy rate of human-object behavior analysis of the method in this paper has been significantly improved on the two databases. |
