Lars 2016 A Texture Driven Approach for Visible Spectrum Fire Detection

A Texture Driven Approach for Visible Spectrum Fire Detection on Mobile RobotsCristiano Steffens, Ricardo Nagel Rodrigues, and Silvia Silva da Costa Botelho

Universidade Federal do Rio Grande – FURG

Centro de Ciências Computacionais

05/01/2023 A Texture Driven Approach for Visible Spectrum Fire Detection on Mobile Robots 2

About

Challenges: Clutter and scene background, Uncontrolled fire flames can assume a variety of

characteristics Can hardly be described using any of the feature

descriptorsthat are widely used for object recognition.

Approach: Color spectroscopy Texture Spatial information.

05/01/2023 A Texture Driven Approach for Visible Spectrum Fire Detection on Mobile Robots 3

A brief overview on the state-of-the-art

Phillips (2002) Chen (2004) Toreyin (2005) Çelik (2007, 2008, 2010) Li (2011, 2012) Kolesov (2010) Mueller (2013)

05/01/2023 A Texture Driven Approach for Visible Spectrum Fire Detection on Mobile Robots 4

A brief overview on the state-of-the-art

Borges (2010) Chenebert (2011)

05/01/2023 A Texture Driven Approach for Visible Spectrum Fire Detection on Mobile Robots 5

DatasetVideos

24 videos 28k frames (51.37% contain fire) 17k annotated regions Creative Commons 3.0 license Variety of fire sources

Uneven illumination Camera movement Different color accuracy settings Clutter Partial Occlusion Motion blur Scale and projection Reflection

05/01/2023 A Texture Driven Approach for Visible Spectrum Fire Detection on Mobile Robots 6

DatasetAnnotations

Rectangle that embraces the whole fire region Very small fire sparkles left out One frame may present zero or many annotations

XML files Each video file has its corresponding annotation file Average flame area is 61512px (~250×250px square) Fire region size/frame size = 8,92%

05/01/2023 A Texture Driven Approach for Visible Spectrum Fire Detection on Mobile Robots 7

Our Proposal

05/01/2023 A Texture Driven Approach for Visible Spectrum Fire Detection on Mobile Robots 8

Random Forests

RFs are o combination of tree classifiers Proposed by Breiman et al. (2001) Attributes are randomly chosen Each tree classifies the sample independently The final class is given by pooling Each tree is built using 2/3 random samples of the

training set

Can deal with many attributes Are easy to understand Have a linear complexity (after training) Each tree can be executed in parallel

05/01/2023 A Texture Driven Approach for Visible Spectrum Fire Detection on Mobile Robots 12

Results

Gren – Only colorBlue – Color + Texture and Temporal

05/01/2023 A Texture Driven Approach for Visible Spectrum Fire Detection on Mobile Robots 13

Results

05/01/2023 A Texture Driven Approach for Visible Spectrum Fire Detection on Mobile Robots 14

Results

05/01/2023 A Texture Driven Approach for Visible Spectrum Fire Detection on Mobile Robots 15

Results

05/01/2023 A Texture Driven Approach for Visible Spectrum Fire Detection on Mobile Robots 16

Results

Cross Validation: Train/ Validation/ Test

MetricsÇelik (2010)

Zhou (2010)

Chenebert (2011) Ours

Recall (TPR) 0.739 0.987 0.990 0.831 Specificity (SPC) 0.317 0.022 0.724 0.988Precision (PPV) 0.654 0.638 0.857 0.982NPV 0.410 0.501 0.979 0.884 FPR 0.682 0.977 0.275 0.012 FDR 0.345 0.361 0.142 0.018 FNR 0.260 0.012 0.009 0.168 Accuracy (ACC) 0.585 0.635 0.890 0.920F1 Score 0.694 0.775 0.919 0.900MCC 0.060 0.036 0.773 0.843

05/01/2023 A Texture Driven Approach for Visible Spectrum Fire Detection on Mobile Robots 17

Thank you!

A Texture Driven Approach for Visible Spectrum Fire Detection on Mobile Robots

[email protected]

Thanks to: