A Study on Lightweight Intersection Pedestrian Detection Method Based on Improved RT-DETR

Model lightweighting is a prominent research focus in object detection. In urban road intersection scenarios, challenges such as large pedestrian populations, significant target scale variations, and intra-class and inter-class occlusions lead to pedestrian detection models with substantial parameter counts and high computational resource demands. This situation conflicts with practical requirements for balancing model size and detection performance. To address these issues, an improved RT-DETR object detection algorithm is proposed, replacing the original ResNet backbone with the lightweight StarNet architecture, thereby significantly reducing model parameters and computational complexity. The incorporation of a FreqFusion module within the model encoder enhances the fusion of high-level and low-level features, improving category consistency and boundary clarity. Additionally, the adoption of the MPDIoU loss function optimizes bounding box regression. Experimental results on the Wider-Person pedestrian detection dataset demonstrate that the proposed algorithm achieves superior detection accuracy compared to conventional methods. Relative to the baseline RT-DETR, the method reduces parameter count by 43.8% and computational complexity by 44.8%, while incurring only a 0.8% decrease in mean average precision, thereby achieving an effective balance between detection performance and model scale.