增强现实技术（Augmented Reality，AR）通过将虚拟物体叠加于真实场景中，实现虚实融合，为使用者创造出富有互动性和沉浸感的体验。与虚拟现实（Virtual Reality，VR）不同，增强现实系统需要考虑真实场景与虚拟物体的遮挡关系，确保虚拟元素与真实环境呈现正确的视觉关系。

本研究首先利用DSO-SLAM算法对真实场景进行三维重建，捕获场景中真实物体的信息及其稀疏矩阵。然后应用光流算法识别运动物体的边缘轮廓，并结合CLAHE算法以减轻光照变化对光流算法的影响。通过引入双阈值筛选来改进Canny算法，有效地剔除了图像中的纹理轮廓，仅保留深度轮廓，减少纹理信息对深度信息的干扰。最后，采用二次优化模型将深度数据均匀地分布到整个图像中，通过比较虚拟物体与真实场景物体的深度值来确定遮挡关系。

本研究对改进的遮挡算法从时间性能和遮挡准确性两个方面进行了详细的比较测试。测试结果表明，本算法在时间性能接近当前先进的AR-Depth算法，满足了实时遮挡系统的需求。在光照条件不足的情况下，本算法在遮挡准确性上表现出了较其他算法更优的性能。本文最后还开发了一个室外大场景下的“丢手绢”教育游戏，通过实际教育案例来验证本文算法的遮挡性能，结果显示本文算法在AR游戏应用中系统延迟低，帧率稳定，无明显遮挡错误，进一步证明了算法的实用性和可靠性。

Augmented Reality (AR) technology realizes the fusion of the virtual and the real by superimposing virtual objects on the real scene, creating an interactive and immersive experience for users. Unlike Virtual Reality (VR), Augmented Reality system needs to consider the occlusion relationship between the real scene and virtual objects to ensure that the virtual elements and the real environment present a correct visual relationship.

This study first employs the DSO-SLAM algorithm to perform 3D reconstruction of real scenes, capturing information about real objects and their sparse matrices. Then, an optical flow algorithm is applied to identify the contours of moving objects, combined with the CLAHE algorithm to mitigate the impact of lighting variations on the optical flow algorithm. After that, by incorporating dual-threshold filtering to improve the Canny algorithm, texture contours in the image are effectively eliminated, retaining only the depth contours, thus reducing the interference of texture information on depth information. Finally, this research adopts a quadratic optimization model to distribute the depth data evenly throughout the image, and determines the occlusion relationship by comparing the depth values of the virtual objects with those of the real scene objects.

This study conducts a detailed comparative test of the improved occlusion algorithm from two aspects: temporal performance and occlusion accuracy. The test results show that the time performance of this algorithm is close to the current advanced AR-Depth algorithm and meets the needs of the real-time occlusion system. In low-light conditions, this algorithm demonstrates superior performance in occlusion accuracy compared to other algorithms. Finally, the occlusion algorithm from this study is applied in an educational case scenario, where a outdoor educational game "Drop the Handkerchief" is developed to validate our occlusion algorithm performance. The results indicate that the algorithm proposed in this paper exhibits low system latency and stable frame rates in AR game applications, without significant occlusion errors. This further validates the practicality and reliability of the algorithm.