Small Target Detection Algorithm for Traffic Signs Based on Improved RT-DETR.

To tackle the issues of low detection accuracy for small traffic signs in Advanced Driver-Assistance Systems (ADAS), we introduce an enhanced model RT-DETR_ASL to make ADAS more accurate and responsive. Firstly, we lighten and optimize the backbone network by substituting the Basic Block with an inverted residual block, thereby reducing the parameter count and enhancing computational speed. Secondly, we integrate a multi-scale deformable attention mechanism into the AIFI feature extraction network, augmenting the recognition and learning capabilities for small targets, which ultimately sharpens the precision of positioning and recognition. Additionally, to bolster the model's performance in detecting small, poorly defined traffic signs, we incorporate the S2 small-target detection layer to refine and strengthen the network's capabilities. During validation, when setting the GIoU (Generalized Intersection over Union) threshold at 0.7, the RT-DETR_ASL model demonstrated a 4.1% increase in mAP50 (mean Average Precision) over the baseline model. Upon further optimizing the loss function, the mAP value soared by an additional 4.51%, surpassing four other mainstream detection methods. Our experiments confirm that the RT-DETR_ASL model significantly enhances the detection accuracy of small traffic signs while maintaining real-time performance, contributing meaningfully to the advancement of autonomous driving assistance systems. It is hoped that the results of this research can make a valuable contribution to the further development of autonomous driving technology. [ABSTRACT FROM AUTHOR]