Abstract:

For the problems of difficult and uncontrollable data acquisition， as well as limited quantity of available rainy day scene samples in the process of unmanned driving perception performance training， a multimodal fusion-based algorithm for constructing rainy day traffic scenes is proposed. Firstly， the rainy day scenes are analyzed and categorized into two models of rain line models and raindrop models for reconstruction. Secondly， a stochastic multisource fusion-based rain line model is proposed， which integrates rain effect from multiple directions and densities. Next， a heterogeneous mapping-based raindrop model is proposed to achieve realistic convex transparency mapping for individual raindrops， coupled with collision prevention design to mitigate cumulative errors of multiple raindrops in the same area. Finally， the two models are integrated to realize reconstruction of rainy day scenes by using various foundational forms. The experimental results show that as rainfall intensity increases， detailed information in the constructed scenes becomes richer initially， with metrics such as entropy and average gradient showing an initial increase followed by a decrease， while image quality continuously decreases， approaching realistic rainy day conditions. With higher rainfall intensity， both interference and detail in the images notably increase， with higher entropy and average gradient， as well as decreased PSNR and SSIM parameters， indicating significant image quality degradation.

Key words: rainy scene samples, rain line model, raindrop model, heterogeneous mapping