A Multi-Scale Multimodal Fusion Method with Attention Mechanism for Constrained Environments

Multi-Temperature environments in industrial intelligent warehouses facilitate the formation of salt-fog, which degrades the performance of both vision and LiDAR sensors. Thus, the salt-fog severely undermines the reliability of key tasks such as object detection and trajectory prediction. To address this challenge, we propose a Multi-Scale Multi-Modal Fusion Method with Attention Mechanism (MMA). The proposed approach consists of two core modules. First, an attention-based fusion module performs interaction between visual and LiDAR features in a unified feature space, realizing information complementarity across modalities. Second, a multi-scale feature enhancement module reconstructs and strengthens the fused features, alleviating local feature loss and semantic degradation caused by salt-fog interference. In addition, an intelligent warehouse simulation environment with controllable salt-fog intensity and dynamic entities is constructed, and the proposed method is systematically evaluated on two representative downstream tasks: object detection and trajectory prediction. Experimental results demonstrate that MMA significantly improves perception performance and prediction stability in varying downstream tasks. Under severe salt-fog conditions, mAP@50:95 for object detection is improved by 2.9%, and the average displacement error for trajectory prediction is reduced by 30% compared to the baseline.