Research On Generative Adversarial Networks-Based Terrain Mapping

Collaborative air-ground robotic system has recently emerged as an important research area that combines the environmental map building and machine learning capabilities to flourish the concept of smart city,in which a collaborative air-ground robotic system can be very powerful in many practical applications.This work aims to use such system to transform the aerial images from UAVs into terrain map exploited by UGVs to perform ground path planning or navigation tasks.However,A lot of previous work heavily rely on large-scale labeled training data which are hard to obtain,since it involves limited flight time and labor labeling costs.To address these problems,The research mainly include three parts:1.In this paper,a convolutional neural network(CNN)classifier is used for terrain mapping,but it depends on a large number of labeled datasets.In order to solve this problem,this paper presents a novel Conditional-GAN-based active terrain mapping(CGAN-ATM)algorithm which integrates Active Learning(AL)strategy into Conditional Generative Adversarial Network(CGAN)framework to build the terrain map efficiently with a very limited number of labeled data.2.We made comparatively deep analysis of our data,the variations of terrain species is quite different.More generally,we illustrated the terrain label as coarse-grained representation(i.e.,grass,pave and concrete),and the variations of the terrain species as fine-grained representation(i.e.the sparsity of grass terrain).The CGAN model can not deal with different coarse and fine grains.Hence,this paper extend CGAN to unsupervised learning fine-Grained representation(FG-CGAN)by introducing the mutual information term.then integrates Active Learning(AL)strategy into(FG-CGAN)framework to build the terrain map more efficiently with a very limited number of labeled data.3.In order to further reduce manual labeling tasks,this paper presents a Hierical Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets Self Guide Terrain Mapping(H-InFoGAN-SGTM),The algorithm learn hierarchical representation of coarse-grained and fine-grained to achieve the goal of specifying different coarse-grained and fine-grained Latent Code to generate different terrain image blocks.Thus Self-Guide generating different terrain image blocks to train convolutional neural network classifier,and it can achieve the same effect as supervised terrain mapping.Finally,in the real scenario,the methods in this paper is verified by relevant experiments.It not only compares the model generate effect,but also compares predicted results of CNN classifier,and finally verifies them with the classical path planning algorithm.Three methods achieve a robust terrain mapping with 90.35%,93.29%,92.37% accuracy respectively,the corresponding average path planning lengths are 1575.4pt,1472.8pt,1480.1pt,respectively.At the end,further research about achieve integrate the laser infomation of UGVs to built a more accurate fusion map.