EFFICIENT ROAD MAPPING VIA INTERACTIVE IMAGE SEGMENTATION
Presenter: Alexander Velizhev
CMRT’09ISPRS Workshop
O. Barinova, R. Shapovalov, S. Sudakov, A. Velizhev, A. Konushin
Introduction
• Roadway monitoring systems are widely-used for supervising road pavement surface and repair planning
Problem statement
• Analysis road pavement only by video sequences
Problem statement (2)
• Object types:– Lane marking – Road patches and defects
• Solution requirements:– High object detection rate – Maximum automation
Problem statement (3)Source image Expected result
Problem details
• Some real examples
Our algorithm outline1. Video rectification
2. Image preprocessing
3. Image segmentation
4. Features calculation
5. Interactive classification
Automaticofflinestage
Interactiveonlinestage
Video rectification
Image preprocessing
Image segmentation
Features calculation
Interactive classification
Video rectification
• Using of raw video has severe drawbacks:– Objects are represented with different
spatial resolution on the same frame– Projective distortions– Elongated objects exceed the bounds of
single frame
Video rectification (2)
Image preprocessing
Image segmentation
Features calculation
Interactive classification
• Video frames are converted to orthogonal projection and stitched to each other
Video rectification
Image preprocessing
Image segmentation
Features calculation
Interactive classification
Video rectification
Retinex transform
Contrast adjustment
Bilateral filter
Image preprocessing
Source image
Image segmentation
Image preprocessing
Image segmentation
Features calculation
Interactive classification
Video rectification
• Main goal is representing all objects of interest as different segments
• We use the hierarchical version of mean shift algorithm
Features calculation
Image preprocessing
Image segmentation
Features calculation
Interactive classification
Video rectification
• More than 100 various features are used for classification of segments
• Feature types:– Colour statistics (colour variance, Lab
components’ percentiles, ... )– Shape statistics (elongation, orientation,
area, …)– Difference with neighborhood of the
segments
Interactive classification
Image preprocessing
Image segmentation
Features calculation
Interactive classification
Video rectification
Interactive classification (2)
Image preprocessing
Image segmentation
Features calculation
Interactive classification
Video rectification
User manually marks object segments
Learning of cascade of classifiers
Automatic classification of the next road part
User corrects classification resultsEnd
Start
Cascade of classifiers
• Cascade of classifiers corresponds image segmentation levels
• We descend a hierarchy from large to small segments and reject segments that do not contain pixels of objects of interest
• Classifier training uses the data passed to a corresponding cascade layer by preceding version of cascade
Why do we use the cascade?
• To solve a problem of unbalanced classes
• To speed-up classification
Online learning
• We introduce an online version of the random forest algorithm
• Special class costing• The algorithm’s code is a part of our
open source “GML Balanced On-line Learning Toolkit ”– http://research.graphicon.ru/machine-learning/gml-
balanced-on-line-learning-toolkit-2.html
Why do we use online learning?
• We don’t need to store all training database in memory
• Short learning time• User actions immediately impact on
the classification results
How to measure system efficiency?• We are modeling “ideal” user actions
to measure the efficiency of the interactive classification
• Efficiency criterion:– a minimal number of mouse’s clicks for
making correct classification
Results
Source imageSegmented
imageAnalysis
result
Results (2)
Image part
Clicks
Manual classification
Interactive classification
Results (3)
Image part
Error,%
Summary
• We present a tool for efficient interactive mapping of road defects and lane marking
• Intensive use of computer vision methods on different stages of our data processing workflow increases usability of the tool
Weak points
• Image segmentation errors can degrade classifier and true object bounds cannot be extracted
• Algorithm is not robust to user mistakes
Future work
• Ultimate goal:Development of the universal semantic segmentation system which can be used for object extraction from large class of images
• Nearest plan:Improving the quality of image segmentation by integration colour and range data
CMRT’09ISPRS Workshop
Efficient road mapping via interactive image segmentation