YOLO-ASE: Enhancing Object Detection With Lightweight Feature Fusion

This paper presents a lightweight pedestrian detection model based on YOLOv11, improved through integrating three novel modules to enhance accuracy and reduce computational costs. First, we introduce Adaptive Kernel Convolution (AKConv), which enables flexible sampling patterns and kernel sizes, significantly boosting the model's adaptability to multi-scale targets. Second, a streamlined neck architecture, SlimNeck, is proposed using GSConv to enhance feature fusion efficiency while reducing parameters and computation. Third, the traditional loss function is replaced with Efficient IoU (EIOU) loss, which optimises bounding box dimensions for faster and more accurate regression. Experiments on the Citypersons dataset demonstrate that, compared to the original YOLOv11, our YOLO-ASE model achieves a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$2.7 \% \text{im}$</tex> provement in accuracy, with 4.7 % fewer parameters and a 6 % reduction in GFLOPs. These enhancements make the model well-suited for real-time pedestrian detection on resourceconstrained edge devices, with potential applications in intelligent surveillance, drone inspection, and autonomous driving.