Traffic signs have country specific standards with defined shapes and specific combination of colours. Despite that, there are 'N' number of factors that make the traffic sign recognition challenging.
« Previously: Solve TSR challenges
Here we discuss in detail the various challenges that we need to overcome to get an efficient traffic sign recognition (TSR) system that is not easily tricked.
One of the important challenges in all image processing algorithms is, to tackle illumination variation. The variation in illumination affects the true colour of the object, thereby reducing the quality of the colour features. Edges are very important for the detection of shapes. High illuminance, low illuminance and shadows affect the contrast of the image.
A pre-processing algorithm, which is invariant to illumination condition of the scene, is required in such a condition. The histogram distribution of luminance channel and statistical properties like 1st, 2nd and 3rd order moments can give insight into the illumination condition of an image. A contrast enhancement algorithm might also help boost up the edges. An illumination-invariant colour model like HSV can be used to maintain the colour information even in bad lighting condition, as shown in figure 2.
__Figure 2: __Two different channels of an image and their corresponding edge images using canny edge detection algorithm (thresholds used are same for both the images) .
The optical axis of the camera is not always aligned to the traffic signs. The difference in alignment imposes a perspective view to the captured images. The curves and uphill-downhill roads also create a perspective change in the scene and deformations are highly expected in such scenarios. Perspective view affects the shape of the object, for example; a circle will look like an ellipse in a different view and a regular hexagon might lose its regularity in a perspective view (see figure 3).
Figure 3: The circular sign in the image looks more elliptical in the image .
Shape feature should be robust enough to handle these deformations due to perspective view. The generalized Hough Transform can handle deformed circles to a great extent. The accumulator array creation and further filtering will accommodate the deformations and disconnected circles too.
Scale invariant features like SIFT and SURF can also be used as a solution to get robust features in case of variation in size. Oriented FAST and Rotated BRIEF (ORB) is another technique that is widely being used as rotation and scale invariant feature. Local Energy based Shape Histogram (LESH) is a scale invariant shape-based feature for object recognition. Deformation invariant features are getting used with a strong classifier to solve the perspective correction problems.
Figure 4: Various deformation invariant features.