M2DGR: A Multi-sensor and Multi-scenario SLAM Dataset for ...

1 [ ] 19 Dec 2021M2 DGR: A Multi-sensor and Multi-scenario SLAM Datasetfor Ground RobotsJie Yin, Ang Li, Tao Li, Wenxian Yu, and Danping Zou ,Abstract We introduce M2 DGR: a novel large-scale datasetcollected by a ground robot with a full sensor-suite including sixfish-eye and one sky-pointing RGB cameras, an infrared camera,an event camera, a Visual-Inertial Sensor (VI-sensor), an inertialmeasurement unit (IMU), a LiDAR, a consumer-grade GlobalNavigation Satellite System (GNSS) receiver and a GNSS-IMUnavigation system with real-time kinematic (RTK) sensors were well-calibrated and synchronized, and theirdata were recorded simultaneously.

2 The ground truth trajectorieswere obtained by the motion capture device, a laser 3D tracker,and an RTK receiver. The Dataset comprises 36 sequences(about 1TB) captured in diverse scenarios including both indoorand outdoor environments. We evaluate state-of-the-art SLAM algorithms on M2 DGR. Results show that existing solutionsperform poorly in some scenarios. For the benefit of the researchcommunity, we make the Dataset and tools public. The webpageof our project : Data Sets for SLAM, Data Sets for Robotic VisionI. INTRODUCTIONI ntelligent ground robots have been emerging in a widerange of applications such as logistics, security, warehouses,cleaning, and food delivery [1].

3 In those applications, therobots need to work reliably in indoor or a mixture ofindoor and outdoor scenes. SLAM (Simultaneous Localizationand Mapping) is the critical enabling technology that allowsthose robots to navigate in those complex scenes, which canconstruct a map of the environment while simultaneouslytracking the location of the robot within the map. ThoughSLAM research has made a substantial progress in the pastdecades [2], [3], existing solutions frequently perform poorlyin practice. For example, visual SLAM may fail at texture-less or dark backgrounds, while LiDAR SLAM could havetrouble with long corridors or open areas.

4 SLAM may alsobecome invalid when the robot takes some unusual actions,for instance, a robot moves into a lift and goes out to the newfloor. Those failure cases motivate us to construct a datasetthatincludes more practical scenarios to facilitate SLAM datasets can speed up breakthroughs and en-able a fair comparison between different algorithms. However,most existing SLAM datasets are designed for autonomousdriving or aerial robots as pointed in [4]. Those datasets arenot the best fit for developing and evaluating algorithms ofground robots. Autonomous cars travel fast on streets androads, while the ground robots move at a much lower speed inremarkably different surroundings, including both indoorandoutdoor scenes.

5 The aerial robots fly freely in 3D space and areAll authors are with Shanghai Key Laboratory of Navigation and LocationBased Services, Shanghai Jiao Tong University. This work was supported byNSFC(62073214). Corresponding Author: Danping Zou quite different from ground robots. Although there areafew datasets targeting ground robots [5], [6], they includeonlya few specific sensors or particular scenes. For logistics robots,catering robots, and service robots, challenging scenarios arefrequently faced, like going into a lift or complete darkness, orgoing from outdoors to indoors. These situations may easilymake existing SLAM methods fail, while they are seldomincluded in existing SLAM benchmark this paper, we introduce a new Dataset for SLAM researchof ground robots, which includes both indoor and outdoorenvironments and contains a rich suite of sensors.

6 The datasetcontains trajectories in highly diverse scenes such as halls,lifts, corridors, and roads. Based on this Dataset , we evaluatethe state-of-the-art SLAM algorithms, including both LiDARSLAM and visual SLAM. The results show that existingSLAM systems perform poorly in at least one situation,indicating further efforts are required to improve the SLAM performance. We summarize major contributions as follows: We collected large-scale sequences for ground robots witha rich sensor suite, which includes six surround-view fish-eye cameras, a sky-pointing fish-eye camera, a VI-sensor,an event camera, an infrared camera, a 32-beam LiDAR,an IMU, and two GNSS receivers.

7 To our knowledge,this is the first SLAM Dataset focusing on ground robotnavigation with such rich sensory information. We recorded trajectories in challenging situations likeentering lifts and complete darkness which are commonlyfaced in practical applications, whereas they are notpresent in previous datasets. We launched a comprehensive benchmark where weevaluated existing state-of-the-art SLAM algorithms ofvarious designs and analyzed their characteristics RELATED WORKA. SLAM with different sensorsGenerally speaking, SLAM can be categorized into vision-based and laser-based ones. Vision-based SLAM, or visualSLAM, can be divided into monocular, binocular, and multi-camera settings according to the number of cameras [3].

8 Though monocular visual SLAM [7] [8] is the most stud-ied topic, it suffers from scale uncertainty and scale visual SLAM [9] [10] takes advantage of knownbaseline distance to calculate metric depth by triangulation,and multi-camera SLAM [11] [12] yields a broader field ofview by using more cameras to enhance the robustness ofSLAM in dynamic environments such as streets. However,it is difficult for pure vision-based SLAM to handle scenes(a) Surround-view fish-eye cameras(b) Sky-pointing camera(c) Forward-looking pinhole camera(d) Event camera(e)Thermal-infrared camera(f) LiDARFig. 1. Our Dataset for ground robots was captured by a rich suite of sensors within various scenarios.

9 Some sensory data are few textures or low illumination conditions. By contrast,LiDAR SLAM is often regarded as a more reliable choice insuch challenges. Nonetheless, LiDAR SLAM also has troublewith long corridors, highly dynamic movements, and , Multi-sensor fusion has been successfully appliedto existing SLAM systems to improve both the accuracy androbustness in practice. For example, ORB-SLAM3 [9] andVINS-Mono [7] integrate vision and IMU. LVI-SAM [13], andR2live [14] tightly integrate vision, LiDAR, and IMU. LIO-SAM [15] loosely integrates LiDAR, IMU, and GNSS. GVINS[16] tightly couples image information, IMU, and further improve the SLAM performance in practice, arecent trend is to explore different sensor some of them as ) Multiple cameras:Multiple cameras have a broaderfield of view than monocular or stereo cameras, which canimprove the robustness in dynamic scenes [17].

10 CoSLAM[18] uses independently moving multiple cameras and gainsthe capability of co-localization and robustness in dynamicmovements. If multiple cameras are mounted on the sameplatform, not only a wider field of view can be achieved,but also the scale ambiguity can be resolved by the baselinesbetween cameras [12]. For example, Multicol-SLAM [11]applies three fixed fish-eye cameras on a helmet, ROVO[19] uses four fish-eye cameras to achieve full coverageof 360 field of views. Panoramic SLAM [20] is anotheropen-source omnidirectional SLAM system, which claims toachieve centimeter-level accuracy even in highly ) Thermal-infrared cameras:Thermal-infrared camerashave gained increasing attention for their perceptual capabilitybeyond the visible spectrum and their robustness regardingen-vironmental changes.