A restoration algorithm prior to detection is needed to correct for the weakening of the sharp edges of the sign board due to motion blur.
In a moving vehicle, the video is being captured at real time. So the movement of the vehicle causes an unavoidable problem- the motion blur. Motion blur weakens the sharp edges of the sign board, and it may seriously affect the detection. Hence, it is necessary to have a restoration algorithm prior to detection.
There are ADAS algorithms which extract stable and sharp edges over frames. Optical flow is one of the well-known algorithms to estimate the motion. Deconvolution with a modelled point spread estimation function provides sharpened images and increases the probability of getting true edges. High level flow diagram for motion restoration is given in Fig 5.
Figure 5: High-level flow diagram of motion restoration.
Frequency domain filtering using estimated motion model is another majorly used motion blur restoration technique. Traffic signs are static, and that really makes the motion estimation problem easier when compared to a deformable moving object.
Most signboards are placed on the sides of the roads and it is very common to see them partially occluded by other objects on the roads (see figure 6). But, since they have regular shapes and symmetry, to an extent, they can be detected.
Figure 6: Example of occluded traffic sign .
Part based object detection methods can handle partial occlusions. Randomised Hough transform is another algorithm for handling broken edges. Local binary pattern, a histogram of oriented gradient etc. with a good classification algorithm, can trace the features and recognise signs even if they are partially occluded. The probability of missing occluded signs is less in the case of predictive tracking based algorithms as it tracks edges over frames. Key points tracking and kernel-based tracking are commonly used techniques for tracking objects.