Landmark based localization : Detection and update of landmarks with uncertainty analysis

Abstract : Mobile mapping is the process of collecting geospatial data with a moving vehicle. These vehicles are often equipped with two types of sensors: remote sensing (cameras, lidar, radar) and geo-localization (GNSS, IMU, odometer). Precise and robust georeferencing has been a major challenge for the implementation of mobile mapping systems. Indeed, in dense urban environments, the masks of signals and multipath errors corrupt the measurements and lead to big positioning errors. High precision IMUs enable to bridge the gaps of positioning and ensure a drift low enough to fulfil the requirements of mapping in terms of accuracy. Nowadays, the hybrid positioning systems (GNSS / IMU / Odometer) are mature enough to provide reliable industrial solutions for the collection of geo-referenced data. National and private mapping agencies have started to collect the required row data for building geospatial repositories at very large scales. However, the very high cost of positioning systems incorporating high precisions IMUs restricts their use to the establishment of geospatial reference data and more affordable positioning solutions are needed for map updating purpose.The objective of this thesis is to provide a low cost positioning solution that can be used on a large number of map updating vehicles.We propose to use one or more cameras on a vehicle as a georeferencing system. Indeed, the vehicle’s trajectory can be estimated using visual odometry techniques. To limit the drift of the trajectory due to the accumulation of errors, we propose a registration on a set of visual landmarks that are precisely georeferenced. These landmarks are reconstructed using the reference data generated by precise and expensive mapping systems. Natural road features such as road markings and traffic signs were chosen as visual landmarks.A local bundle adjustment algorithm has been adapted to estimate the pose of the vehicle using a sequence of images acquired by one or more embedded cameras. A rigorous approach that takes into account the uncertainties enables to tune automatically the weights of every constraint in the equation system of the adjustment and to estimate the uncertainties of the parameters. They are used in a propagation based matching algorithm that accelerates the process of tracking the interest points between the images and eliminate many false matches. This significantly reduces the drift of the visual odometry by reducing the sources of errors. The remaining part of the drift is removed using georeferenced visual landmarks. The process of matching the image sequence with the landmarks is guided by the uncertainty of the poses. It adds a set of absolute constraints in the equation system of bundle adjustment. The drift is drastically reduced. Each step of the algorithm is evaluated on real image sequences with ground truths
Document type :
Theses
Complete list of metadatas

Cited literature [299 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-01586207
Contributor : Abes Star <>
Submitted on : Tuesday, September 12, 2017 - 3:46:05 PM
Last modification on : Wednesday, October 9, 2019 - 4:34:02 PM
Long-term archiving on : Wednesday, December 13, 2017 - 4:06:22 PM

File

TH2017PESC1005_diffusion.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-01586207, version 1

Collections

Citation

Xiaozhi Qu. Landmark based localization : Detection and update of landmarks with uncertainty analysis. Geography. Université Paris-Est, 2017. English. ⟨NNT : 2017PESC1005⟩. ⟨tel-01586207⟩

Share

Metrics

Record views

1079

Files downloads

3480