| With the frequent occurrence of major public safety incidents in China and the severe situation,community fire risk has also become a problem that people have to pay attention to.Therefore,it is urgent to establish a community fire risk assessment and early warning index system.This paper extracts the visual features of community fire risk through image recognition algorithm,and analyzes the risks that may already exist or potentially exist in community pictures or videos combined with natural language processing,so as to effectively prevent community fire accidents.Firstly,a set of evaluation index system applicable to the community fire safety risk screening is designed,and the model considers the interrelationship and weight of factors based on the expert scoring mechanism,which improves the accuracy and reliability of the evaluation results.The evaluation indexes of community fire safety risk contain human factors,environmental factors,facility conditions,firefighting capability and safe evacuation capability,and these discrete and continuous factors are quantified and evaluated through a fuzzy comprehensive evaluation method.The subsequent experimental application can accurately assess the risk level of the application scenario and suggest the related risk hazards.Secondly,this paper constructs an image dataset of community fire safety risks and annotates them accordingly,and extracts the visual feature vectors of the images using the image recognition technique Faster R-CNN,in which the targets of fire safety risk hazards are trained in a targeted manner,and the feature network is selected as Res Net101,in which the highest accuracy rate reaches 94.0%.Third,this paper improves the VL-BERT model to complete the relational perception of fire safety risks in communities.The model adds a masked attention module to help the model better utilize spatial information,and combines the added spatial module to perceive the relationship between different objects to improve the performance of the model.The model uses a generic spatial perception dataset to train the spatial relationship recognition of objects and does comparative experiments on it.Compared with other spatial relationship perception models,the improved VL-BERT model in this paper achieves an average accuracy of 72.3%,which is significantly better than other models.Finally,this paper proposes the FIRE-VLBERT model based on the improved VLBERT model,which takes the fire description feature vector set,the fire term feature vector set and the visual risk feature vector set as inputs to train the model alternately.The experimental results show that the trained FIRE-VLBERT model can sense the fire safety risk of a community and report its fire risk level. |
