Noise-Aware Intermediary Fusion Network for Off-Road Freespace Detection

Freespace detection holds a crucial role in autonomous driving technology, particularly in unstructured offroad scenarios that present additional challenges compared to structured road environments. Multimodal fusion methods are widely recognized as effective strategies for addressing these challenges. However, current fusion methods often overlook the critical issue of noise interference in multimodal data. To address this issue, we propose a novel Noise-Aware Intermediary Fusion Network for off-road freespace detection, named NAIFNet. This framework is specifically designed to mitigate noise interference during multimodal fusion. The key component of NAIFNet is the Noise-Aware Intermediary Interaction (NAII) module, which incorporates a denoising template as an intermediary during the fusion process. The NAII module employs multimodal features as query vectors, while the key and value vectors are derived from the search region features of the denoising template. At the same time, noise-aware interaction ensures effective denoising for data in each modality. Furthermore, in the decoding phase, we introduce the Denoising-Guided Decoder (DGD). Leveraging the denoising template, this decoder achieves more precise feature restoration and effectively mitigates the impact of noise. Extensive experiments on the popular benchmark of the offroad freespace detection dataset (ORFD) demonstrate that the proposed NAIFNet achieves state-of-the-art performance.