To address the problem of local features in point clouds being prone to mismatch between classes during multi-target pose estimation, a robust multi-target pose estimation algorithm based on point-cloud instance segmentation is proposed. First, point-cloud clusters are obtained by segmenting the scene point clouds based on density clustering, and the local feature of the point-cloud clusters are extracted using fast point feature histogram (FPFH) descriptor to describe the local geometry of the point clouds. Then, the random forest classifier is used to classify the aggregated local features of the point-cloud cluster, obtain the category to which the point-cloud cluster belongs, and completes the point-cloud instance segmentation. For each instance in the scene, the features of the scene instance and model are matched using the fast library for approximate nearest neighbors (FLANN) matching algorithm, and the matching points of the points after instance segmentation are obtained on the corresponding category model. Robust initial pose estimation is obtained using random sample consensus (RANSAC) algorithm and the least squares algorithm. Finally, the accurate pose estimation result is obtained using the point-to-plane iterative closet point (ICP) algorithm. The evaluation results in the CV-Lab 3D synthesis and UWA real-scene datasets show that the proposed algorithm significantly improves the interior point probability in the local feature matching stage, thereby improving the robustness and efficiency of pose estimation, particularly in applications with multiple instances in the scene, compared with the direct matching model and local features of all scenes for multi-target pose estimation.