Key points are not available for this paper at this time.
With depth cameras and LiDAR improving and generating more data, their applications in 3D point clouds are growing rapidly. However, the vast amount of generated data increases the computational load and results in a shortage of storage space. Therefore, a preprocessing step to reduce the number of points is required before using the 3D point cloud. This study proposes region of interest (ROI) determination methods that sequentially construct circular and rectangular ROIs along the target trajectory of the robot to extract only crucial data for the target task. These two ROI determination methods have two benefits. First, they maintain the resolution of the raw data; second, they create two ROIs that match perfectly regardless of the complexity of the trajectory. To verify the high performance of these two ROI determination methods, we conducted simulations and experiments using various data; artificial frames, keyframes, and sequential frames. As a result, when the distance between the center points was small, 25% of the diameter or height of the circular and rectangular ROIs, the classification evaluation results were closer to 1 and the processing speed was faster than the raw data acquisition rate. However, we confirm that there is a trade-off relationship between the classification results and the processing time according to the distance parameter. In addition, through the qualitative comparison with the previous study, the long cuboid ROI determination method, we identified the limitations of the previous study and the advantages of the two proposed ROI determination methods.
Park et al. (Sun,) studied this question.