Research And Implementation Of Class-aware Tiny Object Recognition Over Large-scale 3D Point Clouds

Point cloud can be used to represent the appearance of objects in a three-dimensional space and widely used in automatic driving,quality inspection,visualization and ani-mation rendering.As the information infrastructure drives the collaboration of deep learning technology with the cloud and the hardware,people gradually turn their per-spective from two-dimensional vision to three-dimensional vision.Point cloud has be-come preference to study this problem because it can retain the geometric information of objects to the greatest extent.Recently,although tremendous strides have been made in deep learning over large-scale 3D point clouds,one of the remaining open challenges is segmentation of tiny objects.There exist two critical problems hard to be solved:(1)The signal and appearance information of tiny objects are generally sparse and in-sufficient due to their tiny size together with unorderness and permutation invariance,which makes it difficult to be extended to tiny 3D object segmentation with existing studies in semantic segmentation of both tiny 2D object and general 3D object.(2)The sparsity of tiny targets leads to class imbalance.Massive background may dominate the feature learner which makes it extremely difficult to select tiny target samples from the background,resulting in the weak feature representation of tiny objects.To the best of our knowledge,we explore tiny object segmentation for the first time over large-scale3D point clouds by designing deep neural network.Specific to the above two problems,a class-aware tiny object segmentation model is designed and implemented over large-scale 3D point clouds.Furthermore,a real application is selected to explore model's performance:First,a novel two-stage approach,named CTOR,is proposed to explore tiny object segmentation task over large-scale 3D point clouds by designing deep neural network for the first time.It is capable of effectively learning an informative and discriminative feature representation and efficiently processing large-scale point clouds.In the first stage,a class-aware filtering strategy is designed to dispense with the redundant back-ground information.In the second stage,a novel aggregating and sampling scheme is introduced in a supervised manner to progressively increase the receptive field for each 3D point,which can adaptively capture useful signals from tiny objects and learn complex geometric structures.A biased loss function that is inversely correlated to the number of tiny object points is proposed,thereby effectively handling the sparsity of tiny objects.An extensive experimental evaluation on public data set shows the supe-riority of the proposed approach under evaluation metrics.Second,the proposed CTOR solves a real-world application: rock bolt detection.Concretely,rock bolt detection is a fundamental quality inspection task in the mining field.A rock bolt detection system is designed with CTOR embedded,obtaining the accurate position of rock bolts in the tunnel.This is the first attempt to solve rock bolt detection problem by using deep learning technology.Quantitative and qualitative experimental results on private data set verify its effectiveness and superiority.