无人驾驶地图构建和高精度定位论文

【实例简介】
斯坦福大学的博士论文，详细阐述了如何对激光雷达进行标定，如何构建高精度地图，如何用高精度地图进行定位，斯坦福大学无人车的架构和传感器介绍。深入浅出，是学习SLAM和无人驾驶的必读文章。内容是纯英文的，介意勿下。
abstract This dissertation presents several related algorithms that enable important capabilities for self-driving vehicles USing a rotating multi-beam laser rangefinder to sense the world, our vehicle scans mil lions of 3D points every second. Calibrating these sensors plays a crucial role in accurate perception, but manual calibration is unreasonably tedious, and generally inaccurate. As an alternative, we present an unsupervised algorithm for automatically calibrating both the intrinsics and extrinsic of the laser unit from only seconds of driving in an arbitrary and unknown environment. We show that the results are not only vastly easier to obtain than traditional calibration techniques, they are also more accurate a second key challenge in autonomous navigation is reliable localization in the face of uncertainty Using our calibrated sensors, we obtain high resolution infrared reflectivity readings of the world. From these, we build large-scale self-consistent probabilistic laser maps of urban scenes, and show that we can reliably localize a vehicle against these maps to within centimeters, even in dynamic environments by fusing noisy gps and imu readings with the laser in realtime. We also present a localization algorithIn that was used in the DARPA Urban Challenge, which operated without a prerecorded lascr map, and allowed our vehicle to complete the entire six-hour course without a single localization failure Finally, we present a collection of algorithms for the mapping and detection of traf- fic lights in realtime. These methods use a combination of computer-vision techniques and probabilistic approaches to incorporating uncertainty in order to allow our vehicle to reliably ascertain the state of traffic-light-controlled intersections Acknowledgements As members of the Stanford Driving Team, Mike Montemerlo, Dirk Haehnel, Hendrik Dahlkamp, David Stavens, Alex Teichman, Michael Sokolsky, Soeren Kammel, Charles DuHadway, David Jackson, David Held, Ganymed Stanek, Jake askeland, Jan Becker, Jennifer Dolson, J. Zico Kolter, Dirk Langer, Oliver Pink, Christian Plagemann, and moritz Werling contributed to various aspects of this work and to the supporting hardware and software infrastructure used in our autonomous vehicle As fellow students of Professor Thrun, Varun Ganapathi and James Diebel provided valuable insight in conversations about our research Portions of this research were funded by DARA, volkswagen, Google, Intel, Qual- comm, and boeing. I received additional financial support as a National Science foun dation graduate Research Fellow from 2005 to 2008 and from the Qualcomm Innovation Fellowship in 2010 I extend deep thanks to my dissertation advisor, Sebastian Thrun, for his invaluable insight and guidance during this process. Much thanks goes out to the rest of my faculty ommillee: Daphne Koller (Examiner/Reader), Marc Levoy(Examiner/Reader), Vaughan Pratt(reader), and Clifford Nass(examination Chair) I would also like to recognize the support and encouragement of my parents, Arthur and Rita levinson, and my sister, Anya Levinson Contents abstract Acknowledgements 1 Introduction 2 Mapping and localization 2.1 Introduction 2.2 Road Mapping with graphSLaM v155889 2.2.1 Modeling motion 2.2.2 Map Representation 2.2.3 Latent Variable Extension for GPs 2.2.4 The Extended Graph slaM Objective Function 2.2.5 Integrating Out the Map 2.2.6 Computing the Map 13 2.3 Online localization 14 2.3.1 Localization with particle filters 14 2.3.2 Data management 2.4 Experimental Results 18 2.4.1 Mapping 18 2.4.2 Localization 20 2.4.3 Autonomous driving 23 2.5 Conclusion 23 3 Extension to Probabilistic Maps 25 3.1 Introduction 5 3.2 Probabilistic Maps 鲁·鲁 27 3.2.1 Map alignment using GraphSLaM 28 3.2.2 Laser calibration · 28 3.2.3 Map creation 3.3 Online localization 31 3.3.1 Motion update 31 3.3.2 Measurement update 32 3.3.3 Most likely estimate 35 3.4 Experimental Results 35 3.4. 1 Quantitative Result 36 3.4.2 Autonomous success 38 61.5 Conclusion 39 4 Localization USing a Vector Road Map 42 4.1 Introduction 42 4.2 Lane Marker Matching Constraint 46 42.1 RNDF lane marker response prior·.· 47 4.2.2 Laser lane marker response filter 47 4.2.3 Computation of RNdF alignment strength 48 4.3 Curb Avoidance Constraint 49 4.3.1 RNDF lane corridor prior.·.·· 0 4.3.2 Curb response filter 0 4.3.3 Computation of RNDF alignment strength 52 4.4 GPS Constraint 53 4.5 Lateral localization 54 4.6 Results 56 4.6.1 Hard ware requirements 56 4.6.2 Qualitative performance 4.6.3 Quantitative performance 58 4.7 Conclusion 9 5 Unsupervised calibration for Multi-beam Lasers 61 5.1 Introduction 61 5.2 Extrinsic calibration 64 5.3 Intrinsic calibration of each beam 68 5.4 Remittance calibration 5.4.1 Deterministic calibration 70 5.4.2 Bayesian calibration 71 5.5 Experimental Results 74 5.5.1 Extrinsic calibration 74 5.5.2 Intrinsic calibration 77 5.5.3 Remittance calibration 79 5.6 Related extensions 5.6.1 Single-beam laser calibration 81 5.6.2 Calibrating time delays 83 5. 7 Conclusion 84 6 Trafic Light Mapping and state Detection 88 6.1 Introduction 88 6.2 Traffic Light Mapping 91 6.3 Traffic Light State Detection 93 6.3. 1 Prominent failure cases 93 6.3.2 Traffic Light Tracking 95 6.3.3 Uncertainty discussion 97 6.3.4 Probabilistic Template Matching ,,.,99 6.3.5 State Detection Pipeline 102 6.4 Experimental results ..103 6.5 Localization and calibration extension 108 6.5.1 Incorporating Vehicle Localization .108 6.5.2 Extrinsic Camera Calibration 110 6.6 Conclusion l12 7 Conclusions 114 a Hardware and software architecture 117 A 1 Vehicle hardware 117 A2 Software architecture l19 A 3 Object classification 120 A 4 planning ,,,,,,121 A 4.1 Lateral motion 122 A 4.2 Longitudinal movement 122 A 4.3 Combining Lateral and longitudinal curves 123 A. 5 Control l24 Bibliography 127 List of Figures 1.1 Junior was the first robot to cross the finish line in the 2007 darPa Urban Challenge, winning second place overall 2. 1 The acquisition vchiclc is equipped with a tightly intcgrated incrtial navi- gation system which uses GPs, MU, and wheel odometry for localization It also possesses laser range finders for road mapping and localization.... 6 2.2 Visualization of the scanning process: the liDar scanner acquires range data and infrared ground reflectivity. The resulting maps therefore are 3-D infrared images of the ground reflectivity. Notice that lane markings have much higher reflectivity than pavement 2.3 Example of ground plane extraction. Only measurements that coincide with the ground plane are retained all others are discarded (shown in green here). As a result, moving objects such as car(and even parked cars)are not included in the map This makes our approach robust in dynamic envi- ronments 2.4 Patch of the map acquired in bright sunlight on a sunny day(left), and at night in heavy rain(right By correlating scans with the map, instead of taking absolute differences, the weather-related brightness variation has almost no effect on localization 14 2.5 Aerial view of Burlingame, CA. Regions of overlap have been adjusted according to the methods described in this chapter. Maps of this size tend not to fit into main memory, but are swapped in from disk automatically during driving 16 【实例截图】
【核心代码】