In the event of a fire in a logistics center, factory, or office, delays in early detection can lead to failure in initial firefighting, resulting in significant loss of life and property that can derail businesses. To prevent this, fire detection systems are beginning to attract attention for their fast detection speed and accuracy, going beyond sensing and incorporating computer vision and AI.

At the 2023 e4ds AI Tech Concert hosted by e4ds on the 2nd, AMC Planet (hereinafter referred to as AMC) CEO Hyunseok Kim appeared and presented on fire detection AI solutions and algorithms under the theme of 'Utilization of Computer Vision: AI Fire Detection System'.

Traditional fire detection systems have included sensor detection and image analysis. Flame detectors that utilize sensing and have fast and accurate detection rates are widely used on the market, but CEO Kim explained, “They have the disadvantage of having a very limited detection range and being difficult to maintain, and image analysis methods that do not have AI detection have low accuracy and efficiency because they require monitoring personnel on site or are based on pattern matching technology.”

AMC’s fire detection solution, which can detect flames as small as 1cm2 (20 pixels x 20 pixels), the size of a lighter flame, through computer vision, enables immediate detection with a response speed of 0.3 seconds. CEO Kim Hyun-seok emphasized, “It is possible to build a fire detection system with 99% accuracy and low solution cost.”

The main components of the AI fire detection system module architecture are △image preprocessing △neural network △image postprocessing. Preprocessing is the process of processing images so that the neural network can easily understand them. CEO Kim said, “Since ordinary fire scene videos and photos are often deteriorated, noise is removed using Gaussian or median filters.”

The neural network identifies the detection location and type and sends it to image post-processing. Post-processing is responsible for implementing images such as bounding boxes to be shown to the user based on the preceding information. The neural network structure consists of △Backbone △Neck △Head.

CEO Kim explained, “We separately extract features from Backbone, Neck, and Head, extract detailed features by size, and process the bounding boxes of detected objects.” He continued, “This is a method used since YOLO4, and this method can extract features more clearly and prevent the vanishing problem from occurring when the features are passed to the next layer.” Based on this neural network structure, AMC’s solution was able to implement more precise detection and 99% accuracy.

In fire detection solutions, detection functions are important, but in object detection technology, it is also important to implement object tracking well. On this day, CEO Kim explained the implementation of computer vision technology and also mentioned object tracking.

When a video sequence comes in, first we recognize that there is an object (Object Recognition), then we determine what it is (Object Classification), and then we determine its location (Object Localization). The result of these two processes is object recognition, and the task of matching the IDs of each object box by comparing it to the previous frame as a result of object detection is object tracking.

Kim said, “When I observed at the AI competition two years ago, there were many cases where it ended with object detection,” and “processing the bounding box for each frame made it look like tracking the object, but it is not true tracking.” He emphasized that it is necessary to assign an ID to each object and check that there are no ID change/switching issues.

AMC Planet also added separate processing to distinguish between real fire and fake fire during the development process. A dataset of fake fire and real fire images was prepared and a neural network was trained in a separate class.

In addition, the algorithm was designed to distinguish between edges and whether fire is moving between frames during the preprocessing stage so that it can distinguish fake fire even when a mobile phone or camera is held close to the camera.

CEO Kim predicted that in the future, the direction of AI technology development will require research on: △AI utilization in edge devices, △neural network control methods, △methods for reducing neural network training speed and inference time, △response to external attacks, and △neural network security methods using manipulated data.