DD-YOLO: A dual-channel dual-path YOLO network for target detection of blurred vehicles

Abstract

The detection of blurred vehicle targets is essential for maintaining traffic efficiency and ensuring road safety. Although various you-only-look-once (YOLO)-based models exist, few studies have focused on blurred vehicle detection under real-world traffic conditions. To fill this gap, we propose a novel dual-channel dual-path YOLO (DD-YOLO) network, featuring dual-channel feature extraction and dual-path feature fusion. The network comprises a hybrid pooling pyramid (HPP) module, a dual-channel feature extraction backbone, and a dual-path fusion pooling neck (DPFP-neck). Within the DD-YOLO network, we first introduce the HPP module to reduce dependence on key features by combining max and average pooling, incorporating background information to mitigate false positives. Subsequently, the dual-channel backbone is designed to enhance DD-YOLO’s sensitivity for blurred vehicle targets by integrating multiple convolution and attention mechanisms, including the convolutional block attention module (CBAM), simple and parameter-free attention module (SIMAM), standard convolution, and ghost convolution, to capture richer features and improve recall. Finally, the DPFP-neck is developed to fuse diverse information and expand the receptive field across network depths, providing a satisfactory balance between precision and recall. Experiments on the BDD100K and KITTI datasets show that DD-YOLO improves detection accuracy by 4.9% and 4.0%, respectively, with [email protected] gains of 2.4% and 2.7% over the baseline, demonstrating its effectiveness and real-time capability in detecting blurred vehicle targets.