Generating Reference Views of Traffic Intersection for Safe Driving Assistance - Digital Signal Processing for In-Vehicle Systems and Safety - page 195

Fig. 14.1 Reference view
roadside-vehicle communication. With the use of a well-designed human-machine
interface, their system could improve traffic safety to some extent by providing
resulted warning signals to drivers. However, the final information which the driver
receives is that of danger warning. Such information is helpful, but it is not easily
accessible compared to what drivers directly obtain using their eyesight. Also, it is
difficult and awkward to handle at times. In another work that pertains to image
processing, Ichihara et al. extended their NaviView [ 2 ] to suit the environment at the
intersections. But a simple affine transformation can only provide drivers with mirror
image of views from a roadside camera. Though that system could extend the driver's
visual field to the next intersection, it is still power-limited, and the information
obtained is difficult to handle.
We believe that visual information is intuitive. It enhances the driver's ability in
handling the surrounding situation. Note that a driver's visual field is limited by the
vehicle's structure and inter-object occlusion. Broadening it will make it more
efficient. With this in mind, we propose a method for generating a reference view
(Fig. 14.1 ) which follows the vehicle's movement from a higher ground.
The resulted view not only extends the driver's visual field but also provides
information about the vehicle itself. This leads to the strengthening of robustness
against forthcoming occlusions. Since this viewpoint is aligned with the vehicle's
direction, then it has a direct relation with what the driver could see. Also, it is
natural for the driver to handle such view as reference information.
To generate such kind of view, we expect to use roadside cameras located at the
intersections. Nowadays, roadside cameras have been installed in places where
traffic accidents occur frequently, especially at intersections. Using image data
from those cameras will be cost-effective.
We choose image-based rendering (IBR) method to achieve our goal because it
could provide a realistic novel view. Since the shapes in novel view have to be
preserved, the IBR method based on implicit geometry, such as view morphing [ 3 ],
needs to be adopted. The accuracy of these methods has increased in the last decades.
But they have not been widely used in real applications due to their excessive
dependence on manual operations and need for prior knowledge of scene geometry.
In this work, we extend and apply view morphing in a real application by integrating
robust fundamental matrix estimation and feature matching. Our method only requires
a slight adjustment of existing camera settings to make it amenable for practical use.