Mapping and Assessing the State of Traffic InFrastructure

Home  < Partners  < Activities  < Results  < Demos  < Publications  < Workshop 1  < Workshop 2  < Signsteeling  < Hrvatski  <

Results

Detection of traffic signs

• Simple colour-based detection of traffic signs with hardwired thresholds: poor detection (36%-64%), about 6 false positives per frame.
• Geometric-based detection: radial symmetry, Hough transform: insufficient detection (78%-90%), about 20 false positives per frame.
• Detection by a boosted Haar cascade: over 95% correct detection rate, about 1 false positive per frame, 1.6 pixel average localization error normalized to a 24 by 24 detection window [ITSC10].
• Bayesian color-based detection at the pixel level for yellow panels: 80% detection rate at 3% false positive rate has been achieved [ITSC10].
• Enforcement of temporal consistency of traffic sign detections led to better localization accuracy and fewer false positives [submitted] (video).
• Filtering detection responses with a suitably trained strong classifier resulted in further decrease of false positives. When this is combined with the previous technique, the false positive incidence typically falls below 1 in 5000 image frames [CVWW11] (video).

(click to see the video)

Spatio-temporal appearance descriptors

• A novel formalism for describing spatio-temporal appearance has been developed [SCIA11] (video).
  We believe that the developed model shall be useful for detecting objects with sharp occluding borders, such as arbitrary traffic signs, commercial signposts, and other table-like objects.

(click to see the video)

Recognition of traffic signs

• Recognition by cross-correlation: high sensitivity to localization accuracy and motion blur.
• Recognition using pixel-based PCA+1NN, ANN and SVM: around 70% correct recognition rate.
• Recognition by pixel-based LDA+1NN on boosted Haar cascade responses: over 90% correct recognition rate for well-represented classes [ITSC10] (video).
• Recognition by a suitably trained HOG-based SVM tree on boosted Haar cascade responses: over 93% correct recognition rate on individual images, and close to 100% in video [CVWW11].

(click to see the video)

Detection and recognition of road surface markings

• Visual odometry compared to GPS readings.
• Road surface markings are emphasized by the 2nd order Gaussian steerable filter applied to inverse perspective images.
• A method for delimited line detection based on Hough transform achieved 85% correct detection rate on a video of about 2000 images.
• Centerline recognition based on RANSAC-based estimation of the parabolic road model [MIPRO11] (video).

Creating road surface mosaics:

• Limited results produced by a vision-only approach [MIPRO10].
• Encouraging results produced by an approach combining GPS positioning and vision [MIPRO11] (video).

(click to see the video)

Software:

• The first production version of our program for annotating traffic signs in video has been released. Unfortunately, the instructions and user interface are still in Croatian. Please let us know if you would like us to prepare a version in English.

Datasets:

• The three datasets of annotated traffic signs have been released. Please let us know if you use our datasets in your research.