Abstract: Highway abandoned objects endanger traffic safety, easily cause traffic accidents, so it is critical to recognize and clean them up in time. Due to the small area of highway abandoned objects in the image and complex image background, the existing detection methods often have the problems of missed and false detection. To address the above problems, an abandoned objects detection algorithm based on contextual information enhancement and feature refinement is proposed, which is called CIEFRNet. Firstly, a backbone feature extraction module (CSP-COT) incorporating contextual Transformer is designed to fully mine local static and global dynamic context information, and enhance the feature representation of small abandoned objects. In addition, the proposes improved spatial pyramid pooling (ISPP) is used in the backbone, multi-scale downsampling of features is realized by cascade dilated convolution, which reduces the loss of object detail information; in order to improve the feature fusion ability, a feature refine module (CNAB) is designed, in which a proposed mixed attention mechanism (ECSA) is embedded, which can suppress image background noise, and enhances the features of tiny abandoned objects. Finally, it uses the WIoU loss function based on dynamic non-monotonic focus mechanism to improve the learning ability of small abandoned objects and accelerate the network convergence. The experi-
mental results demonstrate that the proposed method achieves 96.5%, 81.6%, 88.1% and 46.5% of accuracy, recall, AP0.5 and AP0.5：0.95 on the self-made highway abandoned objects dataset, respectively, which is better than the currently prevailing object detection methods, and its algorithm complexity is also lower to meet the needs of practical scene applications.

Key words: abandoned objects detection, context information, spatial pyramid pooling, attention mechanism, loss function

摘要： 高速公路抛洒物危及行车安全，极易诱发交通事故，及时识别并清理高速公路抛洒物十分重要。由于高速公路抛洒物在图像中面积占比小且图像背景复杂，现有检测方法常出现漏检和误检的情况。针对上述问题，提出了一种基于上下文信息增强和特征提纯的抛洒物检测算法，记为CIEFRNet。设计了一种融合上下文Transformer的主干特征提取模块（CSP-COT），充分挖掘局部静态上下文信息和全局动态上下文信息，增强小抛洒物的特征表示；主干网络中使用改进的空间金字塔池化（ISPP），通过级联的空洞卷积实现特征的多尺度下采样，减轻目标细节信息的损失；为提高特征融合能力，设计了特征提纯模块（CNAB），其中嵌入了提出的一种混合注意力机制（ECSA），可抑制图像背景噪声，强化微小抛洒物的特征；引入基于动态非单调聚焦机制的WIoU优化损失函数，提高小抛洒物学习能力，加速网络收敛。实验结果表明，所提方法在自制的高速公路抛洒物数据集上的精确率、召回率、AP0.5和AP0.5：0.95分别达到96.5%、81.6%、88.1%和46.5%，优于当前主流的目标检测方法，其算法复杂度也更低，满足实际场景应用需要。

关键词: 抛洒物检测, 上下文信息, 空间金字塔池化, 注意力机制, 损失函数