KGKPD: A Road Object Detection Algorithm that Fuses Knowledge Graphs and Keypoint Detection

To address the challenges of urban traffic complexity, such as occlusion and lighting variations that impact road target detection, this study introduces the KGKPD algorithm. This algorithm integrates knowledge graphs with keypoint detection based on the CenterNet concept. It enhances robustness by introducing salt-and-pepper noise and uses the RepVit network as the backbone. The weighted fusion adaptive feature pyramid network module fuses multi-scale features to optimize the extraction of small target features. The efficient linear deformable convolutional head improves detection of occluded targets. The Poly-1 loss function addresses class imbalance, thereby improving accuracy. The integration of prior knowledge enhances the model’s ability to understand relationships between targets. Compared to CenterNet, the KGKPD algorithm reduces parameters and computational load by 92.32% and 91.85%, respectively, increases the mean average precision by 4.1%, and achieves a frame rate of 40.5 frames per second, meeting the requirements for real-time detection. The code is available at https://github.com/yjx-cup/kgkpd .