Anomaly Analysis is the field of early fire detection through smoke/flame detection in an input image. The Anomaly Analysis Lab is currently monitoring the wildland environment of California and Australia, collecting 2.5 million images per day and is continuously improving the AI performance through the re-learning of both positive and negative detection results.
Wildfire detection at 99% accuracy
Successful wildfire detection from 73,442 images from among a database of 74,146 images of actual wildfire events.
False positive rate (FPR) of 0.0011
Exceptionally low likelihood of detecting non-fire events as wildfire.
Detects early stages of wildfire from a visual input (RGB/IR) of an outdoor wildland environment. The algorithm detects smoke in RGB images and flame in IR images during both day and night.
The AI model is continuously being improved with visual input data that is collected from active CCTV networks located across California and Australia. The application of an AI model creates an inflow of new data that includes error data from both false-negatives and false-positives. The Anomaly Analysis Lab studies not only ways in which errors may be detected automatically via Active Learning, but also ways that let the AI model learn from errors more effectively via Continual Learning.