EVMars-Harsh: An RGB-Event Multimodal Benchmark for Robust Detection in Martian Dust Storm Environments

Event cameras have recently attracted increasing attention in computer vision due to their high temporal resolution, high dynamic range, and low power consumption. However, existing event-based detection benchmarks mainly focus on indoor or real-world driving scenarios, while perception degradation under extreme environments remains insufficiently explored. In planetary exploration and autonomous inspection, severe dust disturbances and complex illumination conditions can degrade visual observations and hinder the spatio-temporal characteristics of event streams, making robust detection challenging. To this end, we introduce EVMars-Harsh, an RGB-Event multimodal detection benchmark for harsh dust environments. EVMars-Harsh is collected in a Mars-like experimental field using a mobile rover platform with synchronized RGB and Color Event sensors, where a controllable blower system is employed to simulate Martian dust disturbances and severe perception degradation. Based on EVMars-Harsh, considering the reduced reliability of event streams under dust interference, we propose a Dust-Aware Sampling module (DAS), which establishes spatio-temporal correlations through scale-adaptive structural continuity analysis to suppress noise events caused by dust particles. Then, we introduce a Reliability-Aware Multimodal Fusion (RAMF) with spatial reliability estimation to adaptively integrate RGB and Event features. Moreover, a Reliability-Aware Lightweight Axial Selective State Space Module (RA-LASSM) is designed to capture long-range dependencies while reducing noise propagation in degraded regions. Experiments on EVMars-Harsh show that the proposed framework achieves 0.743 mAP while maintaining low model complexity, demonstrating its effectiveness for robust RGB-Event detection under severe dust conditions.