Grid-Based Object Tracking

Produktbeschreibung

Mobile robots require an accurate environment perception to plan intelligent maneuvers and avoid collisions. This thesis presents a novel multi sensor environment estimation strategy that fully combines tracking moving objects and mapping the static environment. The basic idea is to fuse and accumulate measurement data by a dynamic occupancy grid model, whereas moving objects are extracted subsequently based on that generic low-level grid representation. Overall, this work results in a robust and consistent estimation of arbitrary objects and obstacles, which is demonstrated in the context of autonomous driving in complex unstructured environments.

Contents
Notations VIII
Abstract XI
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Challenges of Multi-Sensor Environment Perception . . . . . . . . . . . . . 2
1.3 Main Contribution and Outline of This Work . . . . . . . . . . . . . . . . 8
2 Measurement Grid Representation and Fusion 13
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.2 Contribution and Outline . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Evidential Occupancy Grid Representation . . . . . . . . . . . . . . . . . . 16
2.2.1 Spatial Grid Structure . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.2 Evidential Occupancy Representation . . . . . . . . . . . . . . . . . 17
2.3 Sensor Measurement Grids . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3.1 Generic Position-Based Evidential Occupancy Grid Derivation . . . 19
2.3.2 Lidar Measurement Grids . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.3 Radar Measurement Grids . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.4 Camera Measurement Grids . . . . . . . . . . . . . . . . . . . . . . 27
2.4 Measurement Grid Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4.1 Basic Cell-Wise Fusion of Evidence Masses . . . . . . . . . . . . . . 28
2.4.2 Spatiotemporal Alignment of Asynchronous Sensor Data . . . . . . 30
2.5 Results and Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3 Dynamic Grid Mapping and Particle Tracking 39
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.1.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.1.2 Contribution and Outline . . . . . . . . . . . . . . . . . . . . . . . 42
3.2 Dynamic Grid Map and Particle Representation . . . . . . . . . . . . . . . 44
3.2.1 Evidential Frame of Discernment for Dynamic Environments . . . . 44
3.2.2 Dynamic Grid Map Representation . . . . . . . . . . . . . . . . . . 45
3.2.3 Low-Level Particle Representation . . . . . . . . . . . . . . . . . . . 47
3.3 Particle-Based Prediction of the Dynamic Grid Map . . . . . . . . . . . . . 49
3.3.1 Prediction of the Dynamic Evidence Mass . . . . . . . . . . . . . . 50
3.3.2 Prediction of the Non-Dynamic Evidence Masses . . . . . . . . . . 51
3.3.3 Resulting Combined Predicted Dynamic Grid Map . . . . . . . . . 52
3.4 Measurement Update of the Dynamic Grid Map . . . . . . . . . . . . . . . 53
3.4.1 Conflict Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.4.2 Occupancy Differentiation from Distance-Only Measurements . . . 54
3.4.3 Additional Radar- and Camera-Based Occupancy Classification . . 58
3.4.4 Adapted Occupancy Convergence by Object Tracking Feedback . . 62
3.4.5 Overall Resulting Updated Evidence Masses of the Map . . . . . . 63
3.5 Weighting and Resampling of the Particle Population . . . . . . . . . . . . 63
3.5.1 Cell-Wise Occupancy-Based Number of Desired Particles . . . . . . 64
3.5.2 Radar- and Camera-Based Particle Velocity Weighting . . . . . . . 65
3.5.3 Initialization of New Particles . . . . . . . . . . . . . . . . . . . . . 66
3.5.4 Resampling of the Particle Population . . . . . . . . . . . . . . . . 70
3.6 Augmented Measurement Grid . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.7 Results and Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4 Object Extraction and Association 79
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.1.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.1.2 Contribution and Outline . . . . . . . . . . . . . . . . . . . . . . . 81
4.2 Overview of the Extraction and Association Strategies . . . . . . . . . . . 83
4.2.1 Object Detection Based on Dynamic Occupancy Classification . . . 83
4.2.2 Measurement Abstraction Levels of the Association Problem . . . . 84
4.3 Cell Association for Existing Object Tracks . . . . . . . . . . . . . . . . . . 86
4.3.1 Association Based on Predicted High-Level Object Track . . . . . . 86
4.3.2 Particle Labeling Association . . . . . . . . . . . . . . . . . . . . . 87
4.3.3 Additional Clustering with Verification . . . . . . . . . . . . . . . . 91
4.4 Extraction of Newly Occurring Object Tracks . . . . . . . . . . . . . . . . 94
4.4.1 Density-Based Clustering of Dynamic Occupied Cells . . . . . . . . 95
4.4.2 Additional Region Growing with Velocity Variance Analysis . . . . 96
4.5 Results and Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5 Object State Estimation 101
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.1.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.1.2 Contribution and Outline . . . . . . . . . . . . . . . . . . . . . . . 104
5.2 Object State Representation . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.3 Dynamic State Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.3.1 Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.3.2 Transformation of Associated Cells to the Box Representation . . . 108
5.3.3 Position Measurements with Reference Point Selection . . . . . . . 109
5.3.4 Velocity and Orientation Estimation by the Particle Tracking . . . 111
5.3.5 Orientation Estimation Based on Freespace Evidence . . . . . . . . 112
5.4 Additional Radar-Based Dynamic Estimation . . . . . . . . . . . . . . . . 113
5.4.1 Association of Radar Doppler Measurements . . . . . . . . . . . . . 114
5.4.2 Geometric Relations of the Radial Velocity Component . . . . . . . 114
5.4.3 Radar-Based Motion Estimation . . . . . . . . . . . . . . . . . . . . 115
5.5 Shape Estimation and Object Classification . . . . . . . . . . . . . . . . . 119
5.5.1 Histogram Filter Geometry Distribution Estimation . . . . . . . . . 119
5.5.2 Classification Based on Geometry and Velocity Information . . . . . 120
5.5.3 Combined Object Classification with Camera Information . . . . . 123
5.5.4 Extraction of Estimated Length and Width of Box Model . . . . . . 123
5.6 Results and Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
6 Evaluation 127
6.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.1.1 Sensor Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.1.2 Main Processing Steps of this Work . . . . . . . . . . . . . . . . . . 128
6.1.3 Primary Grid Configuration and Algorithm Implementation . . . . 130
6.2 Dynamic Occupancy Grid Estimation . . . . . . . . . . . . . . . . . . . . . 133
6.2.1 Accumulation over Time . . . . . . . . . . . . . . . . . . . . . . . . 133
6.2.2 Comparison with Original Approach . . . . . . . . . . . . . . . . . 135
6.2.3 Occupancy Classification with Additional Information . . . . . . . . 139
6.3 Object Detection and Tracking . . . . . . . . . . . . . . . . . . . . . . . . 147
6.3.1 Object Extraction and Association . . . . . . . . . . . . . . . . . . 148
6.3.2 Dynamic State Estimation for Highly Dynamic Maneuvers . . . . . 152
6.3.3 Object Shape Estimation and Classification . . . . . . . . . . . . . 159
6.4 Summary and Overall Approach Application . . . . . . . . . . . . . . . . . 163
7 Conclusion 165
Own Publications 169
Bibliography 170

Keywords: Autonomes Fahren, Objekterkennung, Objektverfolgung, Occupancy Grid Mapping, Sensordatenassoziation, Sensordatenfusion, Umfelderfassung, Umgebungswahrnehmung, Zustandsschätzung , Autonomous Vehicles, Data Association, Environment Perception, Moving Object Detection, Object State Estimation, Object Tracking, Occupancy Grid Mapping, Sensor Data Fusion