Road detection algorithm

    Our task was to develop the algorithm for the automatic road detection in radar images. The challenge was that the radar images are a bit different from the optical ones. In particular, in the case of synthetic aperture radar (SAR), the image formation process is accomplished via coherent processing of the received signals backscattered from the Earth surface. As a result, the multiplicative speckle noise appears in the SAR images. It complicates the analysis of the image content and extraction of the features of interest. The main idea of the road detection is related with the fact that the road has some principal features. In particular,

– the length of the road is much greater than the width
– the road appear as linear segments with different inclinations in radar images
– the pixel intensity gradients are almost constant along the road, while changing
across the road.

We have used the above ideas in the developed algorithm. Let’s consider an example of SAR image with roads.

Fig.1
Fig. 1. Radar image with roads

We can observe some road segments together with other objects. In order to suppress the noise, we have applied the bilateral filter (https://app.box.com/s/msuvlgq15kn1v%20bkbu7z0xnck5joaw8g4) The filter is based on both spatial and range weighting of the image pixels. The filter weights are determined as

\LARGE \LARGE w(i,j,k,l)=e^{-\frac{(i-k)^2+(j-l)^2}{2\sigma_{d}^{2}}} e^{-\frac{(I(i,j)-I(k,l))^2}{2\sigma_{r}^{2}}}

where the first term corresponds to the spatial weighting with variance \inline \dpi{100} \LARGE \sigma _{d}^{2}  and the second term is related with with the intensity range weighting with variance \inline \dpi{100} \LARGE \sigma _{r}^{2} . Thus, the filtered image can be calculated as follows:

\LARGE I_{b}(i,j) = \sum_{k=1}^{N_{w}}\sum_{l=1}^{N_{w}}I(k,l)w(i,k,l)/\sum_{k=1}^{N_{w}}\sum_{l=1}^{N_{w}}w(i,j,k,l)Fig. 2 shows an example of the bilateral filtering:

bilateral_filter
Fig. 2. Radar image after bilateral filtering (image after filtering and two patches)

After comparison of 2 patches, we can clearly see that road edges are preserved while the high-frequency noise have been suppressed.

In order to catch the road edges we utilized the stroke width transform algorithm (SWT). The original method has been developed for the text detection in natural scenes from optical images (link to the original paper).

In order to construct the SWT image, we need to use the gradient direction image and the corresponding Canny edges map. Fig. 3 illustrates all calculated images:

Fig.3_abc
Fig. 3. Construction of SWT image (a – gradient direction image, b – Canny edge map, c – SWT image)

In order to additionally increase the detection rate, we have integrated the contour detector into the road detection algorithm. A specifically developed image gradient analysis procedure was applied for each of the found contours. Thus, a fusion of SWT and contour detectors provided the expected result.

Fig. 4a

Fig. 4b
Fig. 4. Example of extracted contours and detected road segments

In Fig. 4 top picture illustrates the detected contours. The bottom picture contains the automatically detected road segments.

Recently Ievgen Gorovyi has presented above ideas at the International Radar Symposium (IRS-2015) in Dresden and described how to use the detected roads for the in Dresden and described how to use the detected roads for the estimation of the moving target parameters (https://app.box.com/s/l6n1jar05a48xu9i3wzw1c33g4lstc7f).

Road detector