With advancements in autonomous driving technology, reliable object detection under complex lighting conditions has become a critical challenge. Visible light (RGB) sensors suffer from performance limitations in low-light and backlit scenarios, while infrared (IR) sensors, although capable of compensating for these shortcomings, lack texture details. This paper proposes DMGS-YOLOv8, a dual-modal lightweight object detection framework based on an improved YOLOv8n. The framework employs a parallel dual-stream backbone network to extract features from visible and infrared images separately and to leverage multi-scale deep feature fusion to effectively harness the complementary information from both modalities. To address the increased computational complexity of dual-modal models, we design a lightweight C3 module with Group Shuffle Convolution (C3GS). This module utilizes GSConv as its core component, which significantly reduces computational complexity while maintaining strong feature representation capabilities. Experiments on the LLVIP and M3FD datasets show that, compared to the single-modal YOLOv8n, our DMGS-YOLOv8 achieves an 8.97% higher Average Precision at 50% Intersection over Union (AP50) on the LLVIP dataset, while reducing computational cost (GFLOPs) by 46.67%. It also demonstrates superior overall performance on the M3FD dataset, achieving a balance between detection accuracy and efficiency.