PSCS-Net: Perception Optimized Image Reconstruction Network for Autonomous Driving Systems

International audienceThe progress achieved in transportation systemsand artificial intelligence has amplified the use of intelligenttransportation systems and Autonomous Vehicles (AVs). Indeed,AV systems have attracted much research in recent years, whichenabled multiple autonomous driving tasks, including sceneunderstanding, visual prediction, decision-making, and commu-nication. The latter may create a bottleneck in low-resourceautonomous driving systems that send the collected images toremote edge servers for processing and decision-making. Suchan issue can be addressed by compressing the images in theAV and ensuring a good-quality reconstruction at the edge. Inthis paper, we propose a deep neural network for CompressedSensing (CS) based image reconstruction that integrates imagesemantic perception to improve the reconstruction process forvisual prediction tasks. The reconstruction process is optimizedusing a perception-inspired loss in an end-to-end model learningprocess. The trained model is evaluated on autonomous drivingcar datasets. Obtained experimental results outperform state-of-the-art approaches in terms of both image reconstructionquality and processing time. Finally, we perform semantic urbanscene segmentation on the reconstructed image to evaluatereconstruction quality for visual task prediction. Obtained resultson three semantic urban scene datasets demonstrate the efficiencyof the proposed approach