The paper deals with the analysis of the visual images obtained from fire detection systems. We review the existing approaches to the analysis of video surveillance data and pro-pose a tool for data labeling and visualization. The proposed solution for visual image analysis is based on a neural network (object detection technology). Recognition of hazard locations was carried out using the EfficientDet-D1 model. Video pre- and post-processing algorithms were implemented to improve visual image classification. The pre-processing was used to generate a frame preserving the features of objects that dynamically change over time. The post-processing combines the results of sequential detection of characteristic features on each frame, in particular, features of a smoke cloud. The results of the system operation are presented: visual image classification accuracy was 81%, while localization accuracy was 87%.