Moving object detection (MOD) is a crucial research topic in the field of computer vision, but it faces some challenges such as shadows, illumination, and dynamic background in practical application. In the past few years, the rise of deep learning (DL) has provided fresh ideas to conquer these issues. Inspired by the existing successful deep learning framework, we design a novel pyramid attention-based architecture for MOD. On the one hand, we propose a pyramid attention module to get pivotal target information, and link different layers of knowledge through skip connections. On the other hand, the dilated convolution block (DCB) is dedicated to obtain multi-scale features, which provides sufficient semantic information and geometric details for the network. In this way, contextual resources are closely linked and get more valuable clues. It helps to obtain a precise foreground in the end. Compared with the existing conventional techniques and DL approaches on the benchmark dataset (CDnet2014), the experiments indicate that the performance of our algorithm is better than previous methods.

运动对象检测（MOD）是计算机视觉领域中一个至关重要的研究主题，但在实际应用中它面临一些挑战，例如阴影，照明和动态背景。在过去的几年中，深度学习（DL）的兴起为克服这些问题提供了新的思路。受到现有成功的深度学习框架的启发，我们为MOD设计了一种新颖的基于金字塔注意力的架构。一方面，我们提出了一个金字塔注意模块来获取关键的目标信息，并通过跳过连接来链接不同层次的知识。另一方面，膨胀卷积块（DCB）专用于获得多尺度特征，该特征为网络提供了足够的语义信息和几何细节。这样，上下文资源紧密联系在一起，并获得了更多有价值的线索。最终有助于获得精确的前景。与基准数据集（CDnet2014）上的现有常规技术和DL方法相比，实验表明，我们的算法的性能优于以前的方法。