حسام صفري ترزيلو

Today, the subject of safety and safe driving in autonomous vehicles is of great importance and has attracted a lot of attention in industry and academia. For such intelligent vehicles, the issue of precise lane detection for correct path following is a vital matter. The performance of the systems employed for these purposes is usually affected by a variety of distortions in different forms. Although different solutions and algorithms have been devised for dealing with such artifacts, existing solutions haven’t been able to fully attenuate their destructive effects on the system performance. In this paper, we focus on the stability and robustness of lane detection systems. We present a method with which the accuracy and the performance of a lane detection system remain intact as it encounters new sets of test images with unseen distortions. This is done without imposing a heavy computational load on the model or changing its architecture, but by simply applying the image corruption techniques of different severities on the training datasets and training the models with a combination of these data alongside the main dataset. We eva‎luate the performance of our method on a benchmark consisting of 450 original images from the TuSimple dataset, with 9 kinds of corruption at 3 severity levels per corruption type (12.6K in total). Results showed that our method was able to achieve satisfactory performance in terms of robustness and stability.

رانندگي خودران , تشخيص خط , يادگيري عميق , استحكام , فساد تصاوير

autonomous driving , lane detection , Deep Learning , Robustness , Image Corruptions

Hesam Safari

Abdollah Amirkhani