3D Object Detection Based on Point Cloud Data

Abstract

In the Industrial robotic, computer vision is an important part of the system. The popular object used in the industrial field is a 3D pipe. The problem that is currently being developed is how to detect an object. This research aims to estimate the object detection that is, in this case, is a 3D pipe in various lighting conditions. The camera used in this research is Time of Flight. The methods applied are Remove NaN data for Pre-processing, Random Sample Consensus (RANSAC) for Segmentation, Euclidean Distance for Clustering, and Viewpoint Feature Histogram (VFH) for the object detection. A study conducted on five different objects found that the system could detect each one with a success rate of 100% for the first object, 98.05 percent for the second object, 93.97 percent for the third object, 94 percent for the fourth object, and 99.48 percent for the fifth object. Overall, the system's accuracy in detecting the object is 97.1 percent when four different lighting conditions are applied to five different objects in total.

Density-based Clustering for 3D Stacked Pipe Object Recognition using Directly-given Point Cloud Data on Convolutional Neural Network

“…L. Duc Hanh et al [20] and D. M. Sari et al [3] use Euclidean clustering algorithm [2] to separate all objects captured in the scene into partial clusters of objects. Based on the distance neighborhood of points using Euclidean distance, the algorithm can separate well the object that naturally separated by the distance itself, so the algorithm creates clusters based only on their distance.…”

“…This task cannot be separated from the machine vision system that helps the robot see the environment of the work area. To obtain the visual information, Time-of-Flight camera is one of many 3D sensors or depth camera that provide depth information to represent the geometry and surface of the scene as well as its robustness in facing various lightning [3].…”

“…Also, global descriptors like Viewpoint Feature Histogram (VFH) [11] and Ensemble of Shape Functions (ESF) [12] are commonly used to recognize an object. By using a handcrafted feature extractor, it has a limitation that can't recognize multiple objects in a scene, [3] shows the limitation on recognizing objects is taken one-by-one object which decreases the efficiency of the system's pipeline. Moreover, recently deep learning has dominated many research areas such as speech recognition, natural language processing, and especially computer vision [13], [14].…”