點雲匹配（或稱為點雲套合）是電腦視覺領域中的一項關鍵任務，目的在於將兩個來自相同場景不同掃描角度的點雲場景映射到相同坐標系上，並將兩者對其以還原原始的3D點雲場景，在實務中具有廣泛應用，例如場景重建、自駕車導航以及資料重組等。為此，我們針對現有的點雲匹配方法進行研究，發現多數方法仰賴於對潛在匹配區域的預先猜測，再進行後續的處理與學習。然而，此類方式容易產生過多不必要的匹配假設，導致模型難以聚焦於真正具有意義的重疊區域，進而浪費大量計算資源。此外輸入點雲分別位於不同的座標系下，亦可能在學習過程中產生歧義，使模型難以正確地對齊兩組資料，影響最終的匹配效果。 對於兩個問題我們分別提出了方法進行解決，在預先猜測匹配的部分，我們藉由提取點雲的幾何特徵，然後藉由幾何特徵做出了特徵權重的篩選，將不重要甚至會影響匹配成效的點篩除，而對於兩個座標系的問題，我們結合了Coarse-to-Fine的方式先行將點雲進行初步投射，再利用投射後的結果進行匹配，這樣的方法不論在重疊率高或低的資料集中都取得了良好的成效。;Point cloud registration is a key task in the field of computer vision and has wide applications in practice, such as scene reconstruction, autonomous vehicle navigation, and data reorganization. To this end, we conducted a study on existing point cloud registration methods and found that most methods rely on pre-guessing the potential matching regions before subsequent processing and learning. However, this kind of approach tends to produce too many unnecessary matching hypotheses, making it difficult for the model to focus on truly meaningful overlapping regions, thereby wasting a large amount of computational resources. In addition, since the input point clouds are located in different coordinate systems, it may also cause ambiguity during the learning process, making it difficult for the model to correctly align the two sets of data and thus affecting the final registration result. To solve these two problems, we propose respective solutions. For the part of pre-guessing matching, we extract the geometric features of the point clouds, and then use the geometric features to select feature weights, filtering out points that are unimportant or even detrimental to the matching performance. As for the problem of different coordinate systems, we adopt a coarse-to-fine strategy to first perform an initial projection of the point clouds, and then use the projected results for matching. This method achieves good performance on datasets with both high and low overlap ratios.
