3D Point-cloud Segmentation System Based on AI Model

Abstract

With the rapid development of technology, big data, Internet of Things, and deep learning technologies are gradually developing, among which the development of 2-D image recognition in deep learning is becoming more and more mature. For these reasons, the image recognition of 3D point-cloud is still under development. We combine projection and point-cloud algorithms for large scale semantic segmentation of 3D point-cloud images, and then combine the new neural network based on "Local Flattening for Point Convolution" and "Random Sampling Network" to lighten the model by using depth-separable convolution and quantization. Finally, we implement the convolutional layers in the neural network based on a digital integrated circuit design flow based on a standard cell library.