Pose Estimation Based on Point Pair Features with Optimized Voting and Verification Strategies

In many industrial scenarios, such as automotive production lines, sheet stamping parts are widely used. However, due to the weak texture and strong reflections of these typical parts, pose estimation is hard to achieve, resulting in difficulties of grasping automatically. To deal with this problem, we propose a novel point pair feature (PPF) based pose estimation method to facilitate grasping. Firstly, a three-level structure downsampling method is introduced to seek the balance between the number of model points and significant features. Secondly, in order to reduce the interference of placement plane and other objects in the scene, a two-dimensional voting accumulator is constructed with weighted voting. Based on the voting results, the probability map is accordingly established, which guides keypoints sampling and voting again. Finally, edge points and model points are enrolled for pose verification to remove the wrong results. Our method is implemented in physical experiments, and the results show that the proposed method can be effectively applied to pose estimation of sheet stamping parts such as tire lock plates. Moreover, the ablation study demonstrates the criticality of each process.