Stereo panoramic vision for obstacle detection

Abstract

Statistics show that automotive accidents occur regularly as a result of blind-spots and driver inattentiveness. Such incidents can have a large financial cost associated with them, as well as injury and loss of life. There are several methods currently available to assist drivers in avoiding such collisions. The simplest method is the installation of devices to increase the driver's field of view, such as extra mirrors and wide angle lenses. However, these still rely on an alert human observer and do not completely eliminate blind-spots. Another approach is to use an automated system which utilises sensors such as sonar or radar to gather range information. The range data is processed, and the driver is warned if a collision is immiment. Unfortunately, these systems have low angular resolution and limited sensing volumes. This was the motivation for developing a new method of obstacle detection. In this project, we have designed, built and evaluated a new type of sensor for blind spot monitoring. The stereo panoramic sensor consists of a video camera which views a convex mirrored surface. With the camera and mirror axes aligned, a full 360 degrees can be viewed perpendicular to the sensor axis. Two different mirror profiles were evaluatedthe constant gain, and resolution invariant surfaces. It was found that the constant gain mirror was the most effective for this application. It was shown that the sensor can be used to generate disparity maps from which obstacles can be segmented. This was done by applying the v-disparity algorithm, which has previously not been utilised in panoramic image processing. We found that this method was very powerful for segmenting objects, even the case of extremely noisy data. The average successful obstacle detection rate was found to be around 90%, with a false detecion rate of 8%. Our results indicate that range can be estimated reliably using a stereo panoramic sensor, with excellent angular accuracy in the azimuth direction. In ground truth experiments it was found that the sensor was able to estimate range to within 20cm of the true value, and a maximum angular error of 3°. Through experimentation, we determined that the physical system was approximately half as accurate in comparison to the simulations. However, it should be noted that the system is a prototype which could be developed futher. Nevertheless, this sensor still has the advantage of a much higher angular resolution and larger sensing volume than the driver assistance systems reported to date.