A Compressive Autoencoder For Automotive Photometrical Data

Yasser Almehio, Aymane Hanine, Hafid El Idrissi

Abstract: Automotive lighting systems have been developed as important as most cutting edge systems of new generation vehicles. The new technologies used in headlamps allow to pixelate the scene by light and project many patterns, not only limited to basic lighting beams. This implies a transfer of high amount of data, between the central ECU and the modules, leading to change the standard cost effective networks such as CAN-FD (Controller Area Network with Flexible Data-rate) by Ethernet or LVDS (Low-Voltage Differential-Signaling). In order to achieve optimum data transferring, data must be highly reduced without losing its quality. We introduce a data compression approach that can be applied to all automotive data classes over different transmission mediums. Autoencoders encode images into a latent representation, which is transferred and then decoded to reconstruct the original image. An additional autoencoder can improve lossy data by mitigating compression drawbacks, tested on a dataset with varied lighting features. Our model offers an adaptive tradeoff between image quality and compression rate, achieving a compression rate of 96% alongside a PSNR (Peak Signal-to-Noise Ratio) of 36.4. A booster model is trained too to enhance image quality up to 38 if necessary. Experimental results are compared with JPEG too, and it outperforms in terms of consistent compression ratios and quality preserving in zones submitted to strict automotive regulations. Our approach is lossy, nevertheless, it preserves a global quality necessary to keep photometrical regulations fulfilled. Moreover, the obtained compression rates are apt for transmission via CANFD, a bandwidth-constrained yet highly secure communication protocol prevalent in the automotive industry.