EDP-Net: An Edge-Enhanced Road Object Detection Algorithm with Dual-Path Feature Fusion

Road-object detection for autonomous driving faces challenges including dense occlusions, large-scale variations, and imaging degradation under nighttime and adverse weather conditions, making it difficult for efficient detectors to balance accuracy and robustness. To address this, we propose EDP-Net, an efficient and robust detection network tailored for autonomous driving. First, we design a Learnable Selective Edge Enhancement Module (LSEM) with channel recalibration to suppress background interference while strengthening discriminative edge cues. Second, we construct a Dual-Path Soft-Sparse Feature Pyramid Network (DSFPN), introducing a detail-compensation path to enhance representations of distant small objects. We further develop a deep Global-Local Soft-Sparse Attention Module (GLSAM) to establish long-range context with low computational overhead and incorporate DySample to mitigate cross-scale alignment errors. Finally, we propose a Dynamic Task-Aligned Decoupled Head (DTADH), which employs dynamic task routing and branch-specific alignment to construct task-adaptive features for classification and localization under occlusion and inter-class similarity. Experiments show that EDP-Net achieves 67.0% AP50 and 43.9% AP on BDD-100K. Under a retraining-based evaluation on KITTI, EDP-Net obtains 93.8% AP50 and 74.5% AP, indicating that the proposed architecture remains effective on another autonomous-driving dataset.