<p>To address the challenges of current RGB-D object detection models in road scenes, which are large in size, computationally complex, and difficult to deploy efficiently on resource-constrained devices, this paper proposes a lightweight RGB-D object detection model with adaptive multi-teacher knowledge distillation. First, to construct an efficient detection model, we systematically streamline the dual-stream fusion architecture by simplifying the Cross-Modal Feature Interaction module and designing a Lightweight Cross-Stage Feature Aggregation module that integrates PConv with Triplet Attention. This design effectively fuses multimodal information while significantly reducing model parameters and computational costs. Second, to enhance the detection performance of the lightweight model, we introduce an adaptive knowledge distillation framework driven by multiple teachers. This framework dynamically adjusts the distillation intensity according to sample difficulty and intelligently selects the more competent teacher model for knowledge transfer. Furthermore, to alleviate the scarcity of labeled data, we develop a semi-supervised distillation training strategy that jointly leverages supervised and unsupervised data, thereby effectively improving the model’s generalization ability. Extensive experiments and ablation studies conducted on public datasets demonstrate that the proposed method achieves an excellent balance between model efficiency and detection accuracy, reducing the number of parameters by 54% while achieving mAP@0.5 of 82.2% and mAP@0.5:0.95 of 50.9%.</p>