Domain Adaptation for Infrared-Radar Cross-Scene Multimodal Detection

Utilizing the advantages of both single-mode detections in infrared-radar dual-mode guidance systems is a crucial research direction for precise weapon guidance. Infrared images and radar high resolution range profile (HRRP) have high-dimensional feature spaces, leading to a certain degree of data heterogeneity. The difficulty of capturing targets varies across different scenes, leading to an imbalance in data distribution and consequently, a poor generalization ability of the model. We aim to improve the performance in the target domain by leveraging the knowledge from the source domain and adjusting the model to adapt to the data distribution in the target domain. We propose a dual-mode object detection algorithm, CenterNet-PK, which effectively extracts radar features and establishes temporal correlations using one-dimensional convolution and Bidirectional Gated Recurrent Units (Bi-GRU). Additionally, it incorporates Cross-modal Attention to fully leverage radar features for infrared detection tasks. Building upon this, we develop a domain adaptation (DA) multimodal detection framework to enhance the generalization of the model in various complex scenarios, employing adversarial training to narrow the domain shift. Experimental results demonstrate that the proposed algorithm achieves high detection accuracy and robustness in infrared-radar dual-mode detection tasks. Extensive experiments on DA-detection task from urban to outdoor scenes show that the proposed framework outperforms the baseline. The code is available at https://github.com/chaibosong/CIS-RAM.