DAEAR-DETR: DETR With Dual-Attention and Echo Accumulative Residual for Bus Passenger Detection

Effective detection of bus passengers enhances the intelligence and automation of public transportation systems, but it is challenged by complex backgrounds and severe scale imbalances. To address these challenges, we introduce DAEAR-DETR, a novel neural network architecture employing Dual-Attention mechanisms and Echo Accumulative Residuals (EAR) for bus passenger detection. This model features a Dual-Attention Encoder comprising the Low-Level Local Attention Module (LLLAM) and the High-Level Global Attention Module (HLGAM). Additionally, it integrates a Bidirectional Cross-scale Feature-Fusion Module (BCFM) and a decoder. The Echo Accumulative Residual (EAR) combats information degradation by reintroducing initial input sequences throughout the Transformer encoder layers. A Gating Mechanism (GM) within the EAR connections selectively filters and enhances relevant features based on the ongoing feature process. Experimental results demonstrate the effectiveness of our approach. On the custom bus passenger dataset, DAEAR-DETR achieves 71.9% AP50, 73.5% ARL, and 61.6% AR<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$_{\mathrm {50:95}}$ </tex-math></inline-formula>, outperforming Faster R-CNN, YOLOv9, and other DETR-based methods. Furthermore, DAEAR-DETR demonstrates strong generalization on the public Caltech Pedestrian dataset, achieving Log Average Miss Rate (LAMR) of 4.16% for the Reasonable subset, 4.97% for the Small subset, and 27.79% for the Heavy Occlusion subset. A real-world bus passenger dataset is also created with 15,656 images and 76,006 labeled passengers.