<p>Currently, flame detection algorithms used in firefighting robots are primarily limited to identifying the presence of flames for monitoring and alert purposes. To enhance their utility in firefighting and rescue operations, this paper introduces a real-time flame detection and situation assessment algorithm based on an improved YOLOv8 algorithm. Initially, we defined flame keypoints and designed an algorithm for the automatic annotation of these keypoints to aid in pre-training the algorithm.Subsequently, we improved the network structure of YOLOv8 to enhance flame detection performance, enabling the detection of flames and their keypoints. We further accomplished 3D localization of flame keypoints and a preliminary assessment of the flame situation. Finally, we collected data from the internet, virtual environments, and real environments separately to construct datasets for experiments, validating the effectiveness of the proposed algorithm.Experiments have shown that the accuracy of our flame detection can reach 94.6%, which is 1.5 times that of the original algorithm. The errors in flame width, height, and distance calculated in our fire situation assessment are all within the centimeter level.</p>