本文首發於泡泡機器人SLAM微信公衆號。ICRA ,全名是IEEE International Conference on Robotics and Automation,相信大家都不陌生。今年ICRA2018(5.21-5.25)在美麗的澳大利亞布里斯班舉行,作爲機器人領域的頂級會議,雖然未能親臨現場,但是論文還是要讀的,因此整理了一下與SLAM相關的一些論文,發現SLAM相關的論文確實不少,有70來篇,初略的看了一下,發現今年VIO的文章相對多一些,可能比較關注的是基於DSO的一篇文章,估計有不少同學也在做吧。還有就是與深度學習結合的一些文章,應該也有不少同學關注,大概看了一些深度學習結合,這裏有用來估計位姿,用來做閉環檢測(場景的再識別)等一些思路。還有就是SLAM的一些共性技術,比如優化、數據關聯等,也有不少精彩的論文。這裏按照文章的側重點,對論文進行了一下分類,主要分爲SLAM系統、VO/VIO、語義SLAM/語義地圖、激光SLAM、水下SLAM、拓撲地圖、結合深度學習、場景識別、數據關聯、光度校正、事件相機、場景識別和其他的一些SLAM相關方向。
SLAM系統
今年SLAM系統一共有7篇論文,裏面有的論文是基於地磁做了一個SLAM系統,有的提出了一個輕量級的SLAM系統,,也有針對手持設備提出的SLAM系統,還有一篇論文提出瞭解決運動速度過快,圖像模糊而跟蹤失敗的問題方案,還有orbslam2的一篇論文,最開始還以爲自己看錯了,然後看了下,還真是orbslam2,具體有哪些改進,還沒來得及看。具體的論文如下:
[1] iMag: Accurate and Rapidly Deployable Inertial Magneto-Inductive Localisation
[2] ProSLAM: Graph SLAM from a Programmer’s Perspective
[3] Constructing Category-Specific Models for Monocular Object-SLAM
[4] A Monocular SLAM System Leveraging Structural Regularity in Manhattan World
[5] Complementary Perception for Handheld SLAM
[6] MIS-SLAM: Real-time Large Scale Dense Deformable SLAM System in Minimal Invasive Surgery Based on Heterogeneous Computing
[7] ORB-SLAM2: An Open-Source SLAM System for Monocular, Stereo,and RGB-D Cameras
VO/VIO
這估計是今年比較火熱的一個方向,感覺論文不少,比較關注的可能是基於DSO的imu融合SLAM系統吧!還有一篇以線條來指導關鍵點的選擇,而不作爲特徵的論文,也比較有意思。還有一篇對VIO性能測評的文章,看上去好像也不錯。
[1] Detection and Resolution of Motion Conflict in Visual Inertial Odometry
[2] Direct Sparse Visual-Inertial Odometry using Dynamic Marginalization
[3] Direct Line Guidance Odometry
[4] PIRVS: An Advanced Visual-Inertial SLAM System with Flexible Sensor Fusion and Hardware Co-Design
[5] Dense Planar-Inertial SLAM with Structural Constraints
[6] Low-Drift Visual Odometry in Structured Environments by Decoupling Rotational and Translational Motion
[7] Online Initialization and Automatic Camera-IMU Extrinsic Calibration for Monocular Visual-Inertial SLAM
[8] Relocalization, Global Optimization and Map Merging for Monocular Visual-Inertial SLAM
[9] Towards Globally Consistent Visual-Inertial Collaborative SLAM
[10] Visual odometry using a homography formulation with decoupled rotation and translation estimation using minimal solutions
[11] Encoderless Gimbal Calibration of Dynamic Multi-Camera Clusters
[12] A Benchmark Comparison of Monocular Visual-Inertial Odometry Algorithms for Flying Robots
[13] Robust Stereo Visual Inertial Odometry for Fast Autonomous Flight
語義SLAM / 語義地圖
整體來看,今年語義SLAM的論文好像沒有去年力度大,但是也有一些工作,關注於多幀之間的語義數據融合,還有用語義數據來校正迴環檢測的,我覺得這些工作纔是真正的把語義數據逐漸的利用了起來,而不是像之前的一些論文,直接遞歸貝葉斯,然後加一個CRF搞一下,就說是語義SLAM。具體論文如下:
[1] Multi-view 3D Entangled Forest For Semantic Segmentation and Mapping
[2] A method to segment maps from different modalities using free space layout MAORIS: map of ripples segmentation
[3] Bayesian Scale Estimation for Monocular SLAM Based on Generic Object Detection for Correcting Scale Drift
[4] Semantic Mapping with Omnidirectional Vision
[5] Localization of Classified Objects in SLAM using Nonparametric Statistics and Clustering
激光SLAM
激光SLAM,今年論文好像不是很多,但是關於點雲處理的有一些,3D和2D數據融合的也有一些。
[1] Efficient Continuous-time SLAM for 3D Lidar-based Online Mapping
[2] Elastic LiDAR Fusion: Dense Map-Centric Continuous-Time SLAM
[3] Direct Visual SLAM using Sparse Depth for Camera-LiDAR System
[4] IMLS-SLAM: scan-to-model matching based on 3D data
水下SLAM
之前經常看到有人問有沒有關於水下SLAM的文章,今年ICRA就有,一共有三篇文章,利用水下的聲學數據做SLAM,還有聲吶成像等一些相關技術,具體可以去看論文。
[1] Sonar Visual Inertial SLAM of Underwater Structures
[2] Feature-based SLAM for Imaging Sonar with Under-constrained Landmarks
[3] Pose-graph SLAM using Forward-looking Sonar
拓撲地圖
拓撲地圖作爲導航的一種重要地圖類型,對於SLAM而言還是具有一定的參考價值,因爲對這個也比較感興趣,所以單獨列出來了,大家可以看看。
[1] Topomap: Topological Mapping and Navigation Based on Visual SLAM Maps
[2] Topological Multi-Robot Belief Space Planning in Unknown Environments
[3] Feature-constrained Active Visual SLAM for Mobile Robot Navigation
結合深度學習
深度學習可謂是近幾年來一個非常火熱的領域了,深度學習在圖像復原、圖像分類、分割等領域都取得了突破性進展,但是在2D到3D的視覺研究上面一直也沒有取得突破性的進展。今年ICRA也有很多深度學習相關的論文,這裏主要關注與SLAM結合的領域。今年的深度學習與SLAM結合的領域有:估計位姿、場景的重識別(加語義信息的定位)、對高動態環境自適應等一些領域,具體論文如下:
[1] UnDeepVO: Monocular Visual Odometry through Unsupervised Deep
Learning
[2] Long-term Visual Localization using Semantically Segmented Images
[3] Learning-based Image Enhancement for Visual Odometry in Challenging
HDR Environments
[4] Driven to Distraction: Self-Supervised Distractor Learning for Robust Monocular Visual Odometry in Urban Environments
[6] Addressing Challenging Place Recognition Tasks using Generative Adversarial Networks
[7] Geometric Correspondence Network for Camera Motion Estimation
[8] LS-VO: Learning Dense Optical Subspace for Robust Visual Odometry
Estimation
場景識別
這裏的場景識別,是有兩篇不是基於深度學習的場景識別算法,也可以借鑑一下。
[1] Viewpoint-tolerant Place Recognition combining 2D and 3D information
for UAV navigation
[2] Omnidirectional Multisensory Perception Fusion for Long-Term Place Recognition
數據關聯
數據關聯其實是SLAM系統的一個致命弱點,寧願不關聯,也不能錯誤關聯。我個人覺得,如果我們能夠將前端數據關聯部分和位姿計算部分統一起來,其實是能夠避免這個問題的,因爲誤關聯造成系統崩掉的原因在於,數據關聯和位姿估計是兩個獨立的部分,位姿估計的時候,假設關聯是正確的,如果能夠解決這個問題,我覺得應該可以對這塊有所改善。所以數據關聯其實是一個非常重要的模塊。今年ICRA上也有一些討論的文章,大家可以看一下:
[1] A General Framework for Flexible Multi-Cue Photometric Point Cloud Registration
[2] Pairwise Consistent Measurement Set Maximization for Robust Multi-robot Map Merging
[3] Data-Efficient Decentralized Visual SLAM
[4] Conditional Compatibility Branch and Bound for Feature Cloud Matching
[5] Graph Correspondence-based Point Set Registration
光度校正
光度校正是視覺裏面的一個非常重要的話題,這算是對數據的一個預處理吧,今年視覺領域的best paper提名,其中就有一篇光度校正的文章,大家可以看看:
[1] Online Photometric Calibration of Auto Exposure Video for Realtime Visual Odometry and SLAM
[2] Exposure Control using Bayesian Optimization based on Entropy Weighted Image Gradient
事件相機
新的傳感器的出現,肯定能夠帶動一些SLAM方法的發展,事件相機以及優越的性能受到了研究者的關注,今年ICRA有兩篇論文與此有關:
[1] Ultimate SLAM? Combining Events, Images, and IMU for Robust Visual SLAM in HDR and High Speed Scenarios
[2] The Multi Vehicle Stereo Event Camera Dataset: An Event Camera Dataset for 3D Perception
優化方法
SLAM的優化是SLAM的基礎理論,主要還是關注稀疏性、增量更新、如何消除計算的複雜度、位姿圖等方面。
[1] Complexity Analysis and Efficient Measurement Selection Primitives for High-Rate Graph SLAM
[2] AprilSAM: Real-time Smoothing and Mapping
[3] Fast Nonlinear Approximation of Pose Graph Node Marginalization
[4] A Linear Least Square Initialization Method for 3D Pose Graph Optimization Problem
[5] Convex Relaxations for Pose Graph Optimization with Outliers
[6] BAFS: Bundle Adjustment with Feature Scale Constraints for Enhanced Estimation Accuracy
[7] Robust Incremental SLAM under Constrained Optimization Formulation
[8] Graph SLAM sparsification with populated topologies using factor descent
Optimization
其他
這裏包括一些三角化、定位、特徵編碼、動態場景、迴環檢測等與SLAM相關的技術,具體論文如下:
[1] Robust Visual Localization for Hopping Rovers on Small Bodies
[2] Sparse Gaussian Processes on Matrix Lie Groups: A Unified Framework for Optimizing Continuous-Time Trajectories
[3] Selection and Compression of Local Binary Features for Remote Visual SLAM
[4] Live Structural Modeling using RGB-D SLAM
[5] SLAMBench2: Multi-Objective Head-to-Head Benchmarking for Visual SLAM
[6] Feature-constrained Active Visual SLAM for Mobile Robot Navigation
[7] StaticFusion: Background Reconstruction for Dense RGB-D SLAM in Dynamic Environments
[8] Assigning Visual Words to Places for Loop Closure Detection
[9] Aided Inertial Navigation with Geometric Features: Observability Analysis
[10] Efficient Active SLAM based on Submap Joining, Graph Topology and Convex Optimization
[11] 2D SLAM Correction Prediction in Large Scale Urban Environments
[12] AA-ICP: Iterative Closest Point with Anderson Acceleration
[13] Talk Resource-Efficiently to Me: Optimal Communication Planning for Distributed Loop Closure Detection
[14] Eliminating Scale Drift in Monocular SLAM using Depth from Defocus
[15] Stability-Based Scale Estimation for Monocular SLAM
[16] Ef cient Octree-Based Volumetric SLAM Supporting Signed-Distance and Occupancy Mapping
[17] Incremental Segment-Based Localization in 3D Point Clouds