3d Object Detection And Semantic Segmentation Based On Deep Learning

Compared with two-dimensional images and videos,three-dimensional point cloud data can better reflect the shape characteristics and spatial position relationship of an object,which is of great significance in practical applications,such as autonomous driving,intelligent city planning,home robots,augmented reality and so on.However,compared with semantic segmentation,3D object detection faces more challenges.The increase of dimension makes the amount of computation increase exponentially,which greatly increases the detection and segmentation time.At the same time,due to the more complex 3D space,the accuracy of detection and segmentation cannot reach the twodimensional accuracy.In order to improve the accuracy and efficiency,this paper studies the two tasks of target detection and semantic segmentation,proposes two algorithms and verifies them on related data sets.Firstly,by studying the existing target detection algorithms,a target detection method based on sparse convolution and feature fusion is proposed.Based on the widely used three-dimensional target detection network Voxel Net,this method uses three-dimensional point cloud data as input,uses sparse convolution to extract features,and combines multi-scale information with three-dimensional spatial information through feature fusion network to obtain an efficient three-dimensional target detection network.The experimental results show that this method can effectively detect the target,and the detection speed is 2-4 times higher than the existing methods.Secondly,on the task of semantic segmentation,this paper discusses the advantages and disadvantages of random point sampling and the advantages of feature fusion.On the basis of random point sampling,a 3D semantic segmentation method based on neighborhood expansion and multi-level hierarchical feature fusion is proposed.The receptive field of neural network is enlarged by neighborhood expansion,and more abundant features are extracted by multi-level hierarchical feature fusion.At the same time,the amount of parameters and network depth required for calculation are reduced,and the detection accuracy of semantic segmentation is improved.