SpeedSignRecognition

Speed traffic sign (+ complementary board) detection and recognition algorithm

Uses:

• C++
• OpenCV 3.2

Algorithm:

1. Canny edge detection
2. Ellipse detection with conditions:
   • valid size
   • x, y axis comparison
   • ellipse, contour comparison
3. Ellipse interior rectification (using affine transform)
4. Number detection using a NN + filtering:
   • size filtering
   • gibberish NN result filtering
5. Complementary board detection and rectification under the speed sign:
   • aproximation of contours with polygons
   • finding proper 4 point polygons
   • rectification
6. Cropping of the speed sign and its complementary board(s), printing of detected speed limit

There's also a "backup" procedure doing dilation on Canny edges, which works when the sign is quite close to the camera but the picture quality is bad.

NN:

The NN was trained on digits from The Chars74K dataset by T. de Campos, with 1016 samples per digit (total of 10160 samples). Accuracy on the learning dataset was 99.94 %.

Topology:

• layer 1: 784 neurons (28x28 img size)
• layer 2: 160 neurons
• layer 3: 10 neurons (10 output digits)

Performance:

On a i7 4700MQ, processing for 1 frame requires 9-23 ms for video with downsizing to 1000xN or Nx1000, where N <= 1000.

The real-life speed limit detection accuracy is quite good.

Output example:

input image:

output image:

Detected speed limit: 30