针对当前以目标检测为核心的抛洒物检测算法无法识别“未知类别”的缺陷，以抛洒物引发外观特征变化的视角切入，提出基于特征相似性学习的抛洒物检测方法。首先，在抛洒物体过程中采集参考图像和待检图像，通过参数共享的孪生卷积神经网络得到两张图像的外观特征，然后利用欧式距离等特征相似性函数计算图像区域之间的特征变化并得到欧式距离热力图，最后经阈值筛选得到抛洒物检测结果。为了提升算法对光照等噪声的抗干扰能力，提出全新的注意力掩膜单元，并通过构建长跨度上下文信息和强监督学习的方式提升注意力掩膜的语义判别性能，引导特征响应聚焦于抛洒物引起的外观变化，同时忽略噪声产生的扰动，最终解决噪声干扰和抛洒物产生的特征缠绕问题。为了验证方法的有效性，本研究在真实高速公路场景下进行视频影像数据采集、标注、构建成标准数据集。结果表明：注意力掩膜单元有效提升了特征的语义判别性能，大幅度提高抛洒物检测精度，其中调和均值F_{\mathrm{1}}提高6.4个百分点，同时算法运行速度稳定在30帧/s，满足实时性需求；利用特征序列状态转移方式构建的长跨度上下文信息更有利于注意力掩膜聚焦抛洒物特征信息，抗噪声干扰能力更强； 通过强监督学习得到的注意力掩膜轮廓更为准确，模型精度更高。

To overcome the limitation that the existing dropped object detection methods cannot identify the “unknown category”, this study proposed a dropped object detection architecture based on feature similarity learning. Firstly, the reference image and the query image to be detected were obtained during the dropping process. The appearance features were extracted through a weight-shared siamese convolutional network. Then, Euclidean distance was used to measure dissimilarities between features of reference image and query image. Finally, dropped objects were detected by selecting the pixel from the distance map whose distance value was larger than the fixed threshold. In order to improve its robustness to noise such as illumination change, this paper proposed a novel attention mask unit. And the semantic discriminativeness of the mask was improved through constructing the long-span contextual information and strong supervised learning method. This finally guides the feature response to focus on the appearance changes caused by the dropped objects while ignore the disturbance caused by noise, and solves the problem of feature entanglement between noise and the dropped objects. In order to verify the effectiveness of the method, this study collected data in a real highway scene and built a standard dataset. The results show that the attention mask unit effectively improves the semantic discriminative of features and greatly improves the accuracy of dropped object detection, which achieves F₁ an improvement of 6.4 percentage points. Meanwhile, the algorithm reaches in 30 FPS, which can be performed in real-time. The long-span context information constructed by the feature sequence state transition method is more conducive to attention mask focusing on the projectile feature information, and has stronger anti-noise interference ability. The attention mask contour obtained by strongly supervised learning is more accurate and the model accuracy is higher.