Vehicle Motion Characteristics Constrained Visual Odometry System Research For Intelligent Vehicles

In recent years, visual odometry is one of the most challenging problems in intelligentvehicle research filed, which is also the basis of vision based localization and navigation forintelligent vehicles. Existing methods mostly suffered static scene assumptions,uncertainties, planar motion assumptions, therefore are not suitable for practical applicationfor intelligent vehicles. The topic of this thesis is the visual odometry algorithm forintelligent vehicles moving in complex outdoor environments.For the application in intelligent vehicle system, the on-board visual odometry systemwas modeled, and a co-simulation platform based on vehicle dynamics simulation softwarewas established to analysis the influence of vehicle motion characteristics on theperformance of pose estimation algorithms in visual odometry paradigm.For feature detection and corresponding module, the evaluation metrics for eachsub-module and comprehensive evaluation metric were proposed. Considering the practicalapplication environments, an image dataset was built using our intelligent vehicle platform,and the sub-modules were evaluated and selected based on the image dataset according toproposed evaluation methods. Furthermore the feature detection and corresponding modulewas designed and improved.To solve the problem raised by static scene and planar motion assumptions, two poseestimation algorithms involved vehicle motion characteristics were proposed consideringthe intelligent vehicle running in complex dynamic environments. The first one ismonocular pose estimation algorithm which is suitable for applications requiring vehiclevelocity information: by introducing vehicle dynamic model and reasonable linearapproximation, the relationship between yaw angle variation and vehicle side slip angle wasderived, as well as the pitch angle variation was also involved, then2DoF data associationwas achieved; followed by a two-dimensional histogram voting, the uncertainty ofhypothesis-verify scheme was completely avoided, and the efficient implementation wasensured. The second one is stereo vision pose estimation algorithm which is suitable forapplications that the scene depth information can be obtained. By kinematics modelinitialized3D registration which can convergence efficiently, high-quality inliers werepreserved, therefore the planar constraints was completely relaxed and can be used in harshdriving conditions; and non-iterative refinement was used to ensure the implementation efficiency. Both algorithms take advantage of the motion characteristics of vehicles, bythem the main components of the movement can be captured correctly even abundantdynamic objects exist in the scene, thanks to the ability to eliminate the interference ofoutliers and finally get the right pose estimation.Associated with the feature detection module and pose estimation algorithm module,the complete visual odometry system was integrated. In order to evaluate the accuracy andefficiency of the integrated algorithm, comprehensive experimental study includingsimulations, datasets test and real car experiments were conducted. The system wasvalidated on our co-simulation platform, then tested on foreign datasets of computer vision,and compared with a variety of state-of-the-art algorithms; finally it was validated on realvehicle based actual integration platform. Experiment results on simulations, datasets testand real car test demonstrate that the proposed vehicle motion characteristics based visualodometry system algorithms meet the demand for intelligent vehicle navigation work indynamic and complex environments.