Generative adversarial network inverse perspective mapping image synthesis for autonomous vehicle training / Adizul Ahmad ... [et al.]

Autonomous vehicles (AV) are undergoing extensive research and development due to their disruptive potential and safety advantages. Critical challenges concerning AVs include the precision and accuracy of lane detection. Artificial intelligence (AI) systems for lane detection must be robust, which can be achieved by using many samples to train the system. However, due to the limitations of collection data to account for all road variations is impractical due to the variability of the data involved, especially for highly unique road curvatures. Recent improvements in Generative Adversarial Networks (GANs) make them an attractive tool for generating realistic data to circumvent this problem by generating high-quality images that can represent a variety of road conditions. In this paper, we trained a lightweight GAN architecture on Inverse Perspective Mapping (IPM) images captured by a roof-mounted camera to construct a bird’s eye view (BEV) of the road. After training, GAN was able to generate realistic images with a suitable degree of variation. These variations in the generated images have the potential to train a robust navigation computer aboard the AV to perform curvature estimation. Common testing methods for evaluating GANs are presented (such as SSIM, FID, IS, PSNR) demonstrate that the GAN was able to generate statistically proven realistic images with good variability compared to the original training images. Quantitative results show that images generated using the Exponential Moving Average (EMA) technique achieved a PSNR of 15.7947, SSIM of 0.3875, FID of 131.5848, and IS of 10.5015, indicating improved fidelity and realism over standard outputs.