An Improved Grounding DINO Algorithm for Autonomous Driving

Environmental perception algorithms are a key technology in autonomous driving research, with improving detection accuracy and generalization performance remaining a major focus. However, traditional single-modal object detectors are constrained by their input limitations and often exhibit weak generalization capabilities. While vision-language models enhance generalization through textual modalities, they require manually provided text prompts, significantly limiting their application in autonomous driving. To address this issue, we propose a dynamic text prompt generation strategy (Dyprompt) and contrastive-enhanced attention (CEA) to improve Grounding DINO. Dyprompt automatically generates text prompts from input images, while CEA leverages contrastive matrices to enhance foreground object features. Experiments on the Cityscapes dataset demonstrate that the improved Grounding DINO achieves superior detection performance using only single-image modality, outperforming both single-modal detectors and the baseline model.