Object Detection from 3D Point Cloud Using Deep Learning (SFM)

This paper proposes a cost-effective way for the object detection and classification of objects modeled as 3D renders, via Deep Learning. 3D modeling is the process of manipulating edges, vertices, and polygons in artificial 3D space that creates mathematical coordinated representations of the surface. In this research, we propose to use a stereo camera and a 2D laser scanner (LiDAR) for the construction of 3D object models. We created a 3D model of an object using a stereo camera. Video of objects was captured maintaining the right angles all the time. Then with the help of Intel Real Sense Viewer, a 3D polygon mesh was created, which was converted to a point cloud. A two-dimensional (2D) laser scanner was used to make several chunks of 2D scans from various sides of the object. We then fused the point cloud of the obtained chunks to build a 3D model. We then combined the point clouds obtained from both sources using the Iterative Closest Point (ICP) algorithm. The fused point cloud resulted in the formation of a denser and crispier dataset to be used for Deep Learning. The aforementioned deep learning algorithm, Point Net, encodes sparse point cloud data efficiently and shows very strong performance on par with the state of the art. We have formed a dataset using stereo camera, LIDAR and ICP among which we have obtained the highest accuracy results from ICP algorithm dataset.