Roadcrack Detection Preliminary study by Frederic Maire 1 Marc Miska - - PowerPoint PPT Presentation

f.maire@qut.edu.au 1/15

Roadcrack Detection

Preliminary study by Frederic Maire1 Marc Miska2 Michael Milford1 Sam Dyson1

1School of Electrical Engineering and Computer Science 2School of Civil Engineering and Built Environment

www.roboticvision.

• rg

ARC Centre of Excellence for Robotic Vision

Semantic Vision creating new learning algorithms to detect and recognise

• bjects and places

Vision & Action creating new theory for linking visual percepts and action, for navigation, grasping and human collaboration Robust Vision creating new sensors and robust algorithms to enable operation in real-world situations Algorithms & Architectures creating novel technologies and techniques for computing and communicating.

Research Environment

“Robotics and Autonomous Systems”

www.roboticvision.

• rg

ARC Centre of Excellence for Robotic Vision

First attempt https://www.youtube.com/watch?v=_jIp_kbOB6s

Road crack detection with CUDA

3D plot crack detector

• We have developed a highly discriminative algorithm
• Slow on CPU (~ 1 frame per minute)
• Speed up with CUDA (GPU computing)

implementation on a graphics card

Projective transformation gives a bird eye view of the road

Why is roadcrack detection difficult?

clicked at (581,362) hist values 0.0 0.0 0.0 0.0 0.0 0.0 0.57 0.43 0.0 0.0 0.0

• - - - - - - - - - - - - - - - - - - -

min 6.2 , max 7.7 , (max-min)/avg 0.21 (max-min)/max 0.19

hist values 0.0 0.0 0.0 0.0 0.0 0.011 0.17 0.81 0.011 0.0 0.0

• - - - - - - - - - - - - - - - - - - -

min 5.9 , max 8.0 , (max-min)/avg 0.3 (max-min)/max 0.27

clicked at (413,221) hist values 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.3 0.7 0.0 0.0

• - - - - - - - - - - - - - - - - - - -

min 7.5 , max 8.8 , (max-min)/avg 0.16 (max-min)/max 0.15

clicked at (659,243) hist values 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.95 0.045 0.0 0.0

• - - - - - - - - - - - - - - - - - - -

min 7.1 , max 8.1 , (max-min)/avg 0.13 (max-min)/max 0.12

f.maire@qut.edu.au 11/15

Technical challenges

• Evaluating the “ring distance histogram” for each

pixel is computationally expensive

• Possible approaches

– Use a pre-processing step with adaptive thresholding

and morphological operations to produce a reduced number of candidate roadcrack pixels

– Collect training data, label it using the computationally

expensive algorithm, and train a deep neural network to classify the pixels of new images

f.maire@qut.edu.au 12/15

Deep neural networks

• Standard Convolutional Neural Networks

f.maire@qut.edu.au 13/15

Conversion of a classification net into output a heatmap

f.maire@qut.edu.au 14/15

Conclusion

• Roadcrack detection is a low-hanging fruit!

15/15