Fire and Smoke Detection at AI edgy box
Our AI model, which possesses 1.7 million fire and smoke image data, was trained on a selected 200,000 images. It classifies into 11 different categories, including false positives due to scenes like fireworks and smoke.
It demonstrates a high performance with an average precision (mAP) of 0.90019. Operating on a Jetson Nano, it achieves recognition speeds of 76-85ms per frame, which is 1/20th the performance of a standard AI PC.
It boasts a flame detection rate of over 0.3% of the screen size and can detect as low as 0.1% for clean flames.
Additionally, it offers cross-platform compatibility, supporting Windows, Linux, and ARM architectures.
| ITEM | Spec | Remarks |
| Processor | •128 Core NIVIDA Maxwell GPU •Quad core ARM A57 CPU •4GB 64bit LPDDR4 | |
| Network | 10/100/1000Base-T Ethernet | |
| Power | DC5V, 4A | |
| I/O | •USB 3.0 Type A, 4Ea •USB2.0 micro–B, 4Ea •HDMI/Display Port •Gigabit Ethernet •GPIO, I2C, SPI, UART | |
| SD card | Micro SD | |
| Video Encoder | 4K@30 | 4x1080P@30 | 9x720p@30(H.264/H.265) | |
| Video Decoder | 4K@60 | 2x4K@30| 8x1080P30 | 18x720p@30(H.264/H.265) | |
| AI Algorithm | Smoke & fire | |
| IP Camera | RTSP | |
| Class of Detection | •Black smoke occurrence •Gray smoke occurrence •White smoke occurrence •Fire(flame) occurrence •Cloud •Fog(smoke) •Lighting •Sunlight •Wobbly white object •Leaves, grass, etc, shaken by the wind •Non(Irrelevant) | |
| Recognition Speed | 76~85ms/frame(About 11~13FPS) | |
| Fire frame detection | More than 0.3% of screen size(85x85pixels at 1920×1080 resolutions) | |
| F1-Score | 0.861952 | |
| Dimension mm(Board) | TBD |