计算机科学
计算机视觉
人工智能
目标检测
特征匹配
棱锥(几何)
背景(考古学)
算法
特征(语言学)
对象(语法)
亮度
特征提取
对比度(视觉)
模式识别(心理学)
匹配(统计)
变压器
融合
抓住
内存占用
钥匙(锁)
运动检测
实时计算
精确性和召回率
作者
X X Wang,Zhuorang Yang,Xiang Zhang
标识
DOI:10.1088/1361-6501/ae6d24
摘要
Abstract To address the degradation of object contrast and feature representations caused by low illumination, uneven lighting distribution, and motion blur in nighttime road scenarios, this study presents a lightweight framework for nighttime road object detection. First, Zero-DCE++ is employed to perform adaptive brightness and contrast enhancement on raw nighttime images, thereby improving the detectability of weak-feature objects. On this basis, a lightweight detection network termed SML-DEIM is constructed based on DETR with Improved Matching (DEIM). Specifically, an SDCNet backbone is designed to enhance weak edges and local structural information while expanding the receptive field, a proposed MGD-Mamba module is used to jointly model fine-grained local features and global semantic information through global–local context modeling, and a lightweight multi-scale fusion pyramid network neck is developed to strengthen information interaction and complementary fusion across multiple scales. Experimental results show that the proposed low-light enhancement strategy improves the average mAP@50 by 5.1 percentage points over the baseline model. SML-DEIM achieves the best performance in terms of Recall and mAP, with gains of 3.2 and 3.6 percentage points in mAP@50 and mAP@50:95, respectively, while introducing only limited computational and parameter overhead. These results demonstrate that the proposed method remains lightweight and provides an efficient and practical solution for real-world nighttime road object detection.
科研通智能强力驱动
Strongly Powered by AbleSci AI