To address issues such as difficult pedestrian target recognition, and low detection accuracy in road images for autonomous driving, this paper proposes a road object detection algorithm based on the improved YOLOv11. First, aiming at the limitation of YOLOv11 in multi-scale feature modeling capability, an improved module integrating frequency-domain convolution modeling and spatial-channel attention mechanism is designed. On the basis of the C3k2 structure of YOLOv11, the optimized FDConv module is introduced to construct Bottleneck_FDConv, C3k_FDConv, and C3k2_FDConv sub-structures, which effectively enhance the model's feature extraction ability for small targets and complex scenes. Meanwhile, this paper optimizes the original loss function of YOLOv11: on one hand, Focal Loss is introduced to alleviate the problem of imbalanced positive and negative samples; on the other hand, a small target weighting mechanism is added to improve the model's attention to small targets such as pedestrians. Experiments are conducted on the public KITTI dataset for verification. The results show that the improved YOLOv11 achieves 74.43% in the pedestrian detection metric, which is 3.1% higher than the original YOLOv11. In terms of the mAP50 metric, the improved YOLOv11 algorithm reaches 84.87%, which is 0.85% higher than the original YOLOv11. These results demonstrate that the improved algorithm significantly enhances the model's detection accuracy and robustness, and has high application value.