EMSA-Net: An Efficient Multi-Scale and Boundary-Aware Network for Spillage Risk Vehicle Detection Under Low-Light Urban Conditions
Year: 2025
Type: conference-paper
Abstract: Cargo spillage from improperly loaded or covered trucks, here termed Spillage Risk Vehicles (SRVs), is a significant contributor to traffic accidents. The detection of SRVs on urban roads is crucial for proactive traffic safety management, yet this task is confronted by drastic variations in target scale and severe boundary ambiguity, especially under low-light nighttime conditions. Existing detectors often struggle to balance accuracy and efficiency for this specialized task. To address these challenges, we propose EMSA-Net : an Efficient Multi-scale and Boundary-Aware Network. Our framework introduces two novel and synergistic components: (1) an Efficient Multi-Scale Convolution (EMSC) module, which extracts rich features with minimal computational overhead; and (2) a Selective Boundary Aggregation (SBA) module, which adaptively fuses high-resolution boundary cues with strong semantic information. This dual-attention fusion significantly enhances localization accuracy for poorly illuminated targets. Extensive experiments demonstrate that EMSA-Net significantly outperforms existing state-of-the-art models. Specifically, on UrbanTruck-24H dataset, EMSA-Net achieves 71.8% [email protected] with 92 FPS, outperforming the strong YOLOv8-m baseline by 3.2%. Visualizations further corroborate the model's effectiveness and interpretability, showcasing its capability in robust and precise Spillage Risk Vehicle Detection.
Authors: Ze Xu, Wenting Jia, Chunyang Yan, Ruikang Wang, Yongfu He
Institution: Chongqing Jiaotong University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Awards: G3226253135, G4135145446
OpenAlex ID: W4414396226