Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Traffic Signs Detection / Recognition
-Proposal and Preliminary Results
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Overview
Single-viewSegmentation - extremely fast bounding box selection process with FN → 0.
Traffic signs are designed to be well distinguishable from background ⇒ have distinctive colors and shapes.
Detection - AdaBoost classifier of bounding boxes. Recognition - SVM.
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Overview
Single-viewSegmentation - extremely fast bounding box selection process with FN → 0.
Traffic signs are designed to be well distinguishable from background ⇒ have distinctive colors and shapes.
Detection - AdaBoost classifier of bounding boxes. Recognition - SVM.
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Extension to 3D - Example
Calibrated view 1: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraints Calibrated view 2: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraintsTraffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Extension to 3D - Example
Calibrated view 1: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraints Calibrated view 2: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraintsTraffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Extension to 3D - Example
Calibrated view 1: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraints Calibrated view 2: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraintsTraffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Extension to 3D - Example
Calibrated view 1: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraints Calibrated view 2: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraintsTraffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Extension to 3D - Example
Calibrated view 1: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraints Calibrated view 2: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraintsTraffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Thresholding
Segmentation is often detection of connected components of a thresholded image (T = [0.5, 0.2, −0.4, 1.0]>)
Original Thresholded Connected Segmented
image image I (T ) components bound. boxes
Threshold in RGB space: T = [t, a, b, c]> Thresholded image I (T) = ( 1 a· IR +b· IG +c· IB ≤t 0 otherwise
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Thresholding
Segmentation is often detection of connected components of a thresholded image (T = [0.5, 0.2, −0.4, 1.0]>)
Original Thresholded Connected Segmented
image image I (T ) components bound. boxes
Threshold in RGB space: T = [t, a, b, c]> Thresholded image I (T) = ( 1 a· IR +b· IG +c· IB ≤t 0 otherwise
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Learning segmentation
One threshold insufficient ⇒ set of thresholds T = {T1, T2, . . . } used.
Different T different errors:
FN(T ) ... False Negative (missed traffic signs) FP(T ) ... False Positive (segmented backgrounds) card(T ) ... Cardinality (size of T )
Learning segmentation is searching for T ⊆ P{T} minimizing an error subject to some constraints. Learning is NP-complete (set covering problem) but tractable even for very large P(T).
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Learning segmentation
One threshold insufficient ⇒ set of thresholds T = {T1, T2, . . . } used.
Different T different errors:
FN(T ) ... False Negative (missed traffic signs) FP(T ) ... False Positive (segmented backgrounds) card(T ) ... Cardinality (size of T )
Learning segmentation is searching for T ⊆ P{T} minimizing an error subject to some constraints. Learning is NP-complete (set covering problem) but tractable even for very large P(T).
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Learning segmentation
One threshold insufficient ⇒ set of thresholds T = {T1, T2, . . . } used.
Different T different errors:
FN(T ) ... False Negative (missed traffic signs) FP(T ) ... False Positive (segmented backgrounds) card(T ) ... Cardinality (size of T )
Learning segmentation is searching for T ⊆ P{T} minimizing an error subject to some constraints. Learning is NP-complete (set covering problem) but tractable even for very large P(T).
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Experiments - Segmentation
Neyman-Pearson ( = 5%, card(T) = 104 ⇒ card(P(T)) = 2104 )Boolean linear programming optimization time less 2 sec. card(T ) ≈ 30 thresholds.
FP ≈ 103 bounding boxes per 1236 × 1628 image.
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Feature normalization
Haar featuresFeature normalization (RGB, HSL, HSV, gray) Feature selection (greedy construction by AdaBoost)
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Feature normalization
Haar featuresFeature normalization (RGB, HSL, HSV, gray) Feature selection (greedy construction by AdaBoost)
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
AdaBoost
AdaBoost results: 154 features selected from 5 · 104features. 77 minutes learning on 2400 training samples.Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
AdaBoost
AdaBoost results: 154 features selected from 5 · 104features. 77 minutes learning on 2400 training samples. 0 10 20 30 40 50 0 10 20 30 40 50 FP[%] FN[%] RGBTraffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
(AdaBoost) Cascade
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
(AdaBoost) Cascade
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
(AdaBoost) Cascade
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Multi-class detection by shared cascade [Zehnder et
al. BMVC 2008]
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Multi-class detection by shared cascade [Zehnder et
al. BMVC 2008]
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Multi-class detection by shared cascade [Zehnder et
al. BMVC 2008]
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Extension to 3D - Example
Calibrated view 1: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraints Calibrated view 2: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraintsTraffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Extension to 3D - Example
Calibrated view 1: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraints Calibrated view 2: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraintsTraffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Extension to 3D - Example
Calibrated view 1: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraints Calibrated view 2: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraintsTraffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Extension to 3D - Example
Calibrated view 1: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraints Calibrated view 2: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraintsTraffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Extension to 3D - Example
Calibrated view 1: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraints Calibrated view 2: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraintsTraffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Extension to 3D - Example
Calibrated view 1: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraints Calibrated view 2: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraintsTraffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Extension to 3D - Example
Calibrated view 1: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraints Calibrated view 2: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraintsTraffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Extension to 3D - Example
Calibrated view 1: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraints Calibrated view 2: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraintsTraffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Extension to 3D - Example
Calibrated view 1: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraints Calibrated view 2: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraintsTraffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Extension to 3D - Example
Calibrated view 1: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraints Calibrated view 2: Segmentation (FP ≈ 103) Detection (FP ≈ 101) Geometry constraintsTraffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Conclusions
Segmentation based on optimal selection of thresholds in color space given a set of constraints in performance.
Detection aimed to be solved with AdaBoost cascades1.
Shared features/cascades2 and post-optimization of the
classifiers3 are directions of work.
Recognition - SVM is in block start.
1
Viola and Jones, CVPR’01 2
Zehnder et al., BMVC’08, and Torralba et al., CVPR’04 3
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Conclusions
Segmentation based on optimal selection of thresholds in color space given a set of constraints in performance.
Detection aimed to be solved with AdaBoost cascades1.
Shared features/cascades2 and post-optimization of the
classifiers3 are directions of work.
Recognition - SVM is in block start.
1
Viola and Jones, CVPR’01
2
Zehnder et al., BMVC’08, and Torralba et al., CVPR’04 3
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Conclusions
Segmentation based on optimal selection of thresholds in color space given a set of constraints in performance.
Detection aimed to be solved with AdaBoost cascades1.
Shared features/cascades2 and post-optimization of the
classifiers3 are directions of work. Recognition - SVM is in block start.
1
Viola and Jones, CVPR’01 2
Zehnder et al., BMVC’08, and Torralba et al., CVPR’04 3
Traffic Signs Detection / Recognition -Proposal and Preliminary Results Introduction Segmentation Detection Conclusions
Conclusions
Segmentation based on optimal selection of thresholds in color space given a set of constraints in performance.
Detection aimed to be solved with AdaBoost cascades1.
Shared features/cascades2 and post-optimization of the
classifiers3 are directions of work. Recognition - SVM is in block start.
1
Viola and Jones, CVPR’01 2
Zehnder et al., BMVC’08, and Torralba et al., CVPR’04 3