Embedded Systems

S2S-Net: Addressing the Domain Gap of Heterogeneous Sensor Systems in LiDAR-Based Collective Perception

by Sven Teufel, Jörg Gamerdinger, and Oliver Bring­mann
In arxiv preprint, 2025.

Ab­stract

Col­lec­tive Per­cep­tion (CP) has emerged as a promis­ing ap­proach to over­come the lim­i­ta­tions of in­di­vid­ual per­cep­tion in the con­text of au­tonomous dri­ving. Var­i­ous ap­proaches have been pro­posed to re­al­ize col­lec­tive per­cep­tion; how­ever, the Sen­sor2Sen­sor do­main gap that arises from the uti­liza­tion of dif­fer­ent sen­sor sys­tems in Con­nected and Au­to­mated Ve­hi­cles (CAVs) re­mains mostly un­ad­dressed. This is pri­mar­ily due to the paucity of datasets con­tain­ing het­ero­ge­neous sen­sor se­tups among the CAVs. The re­cently re­leased SCOPE datasets ad­dress this issue by pro­vid­ing data from three dif­fer­ent LiDAR sen­sors for each CAV. This study is the first to tackle the Sen­sor2Sen­sor do­main gap in ve­hi­cle to ve­hi­cle (V2V) col­lec­tive per­cep­tion. First, we pre­sent our sen­sor-do­main ro­bust ar­chi­tec­ture S2S-Net. Then an in-depth analy­sis of the Sen­sor2Sen­sor do­main adap­ta­tion ca­pa­bil­i­ties of S2S-Net on the SCOPE dataset is con­ducted. S2S-Net demon­strates the ca­pa­bil­ity to main­tain very high per­for­mance in un­seen sen­sor do­mains and achieved state-of-the-art re­sults on the SCOPE dataset.