Bahman Soheilian

Vehicle localization using mono-camera and geo-referenced traffic signs

Vision based localization is a cost effective method for indoor and outdoor application. However, it has drift problem if none global optimization is used. We proposed a geo-referenced traffic sign based localization method, which integrated the constraints of 3D traffic signs with local bundle adjustment to reduce the drift. Comparing to global bundle adjustment, Local Bundle Adjustment(LBA) has low computational cost but suffers the drift problem for large scale localization because of the random error accumulation. We reduced the drift by means of the constraints from geo-referenced traffic signs for bundle adjustment process. The original LBA model was extended for the constraints and the traffic signs were detected in images and matched with 3D landmark database automatically. From the experiments of simulated and real images, our approach can reduce the drift and have better locating results than none-constraint LBA based localization method.

Road side detection and reconstruction using LIDAR sensor

Road edge localization is key knowledge for automatic road modeling and hence, in the field of autonomous vehicles. In this paper, we investigate the case of road border detection using LIDAR data. The aim is to propose a system recognizing curbs and curb ramps and to reconstruct the missing information in case of occlusion. A prediction/estimation process (inspired by Kalman filter models) has been analyzed. The map of angle deviation to ground normal is considered as a feature set, helping to characterize efficiently curbs while curb ramps and occluded curbs have been handled with the proposed model. Such a method may be used for both road map modeling and driver-assistance systems. A user interface scheme has also been described, providing an effective tool for semi-automatic processing of a large amount of data.

Augmenting Vehicle Localization Accuracy with Cameras and 3D Road Infrastructure Database

Accurate and continuous vehicle localization in urban environments has been an important research problem in recent years. In this paper, we propose a landmark based localization method using road signs and road markings. The principle is to associate the online detections from onboard cameras with the landmarks in a pre-generated road infrastructure database, then to adjust the raw vehicle pose predicted by the inertial sensors. This method was evaluated with data sequences acquired in urban streets. The results prove the contribution of road signs and road markings for reducing the trajectory drift as absolute control points.

Cross-domain image localization by adaptive feature fusion

We address the problem of cross-domain image localization, i.e., the ability of estimating the pose of a landmark from visual content acquired under various conditions, such as old photographs, paintings, photos taken at a particular season, etc. We explore a 2D approach where the pose is estimated from geo-localized reference images that visually match the query image. This work focuses on the retrieval of similar images, which is a challenging task for images across different domains. We propose a Content-Based Image Retrieval (CBIR) framework that adaptively combines multiple image descriptions. A regression model is used to select the best feature combinations according to their spatial complementarity, globally for a whole dataset as well as adaptively for each given image. The framework is evaluated on different datasets and the experiments prove its advantage over classical retrieval approaches.

Landmark based localization: LBA refinement using MCMC-optimized projections of RJMCMC-extracted road marks

Precise localization in dense urban areas is a challenging task for both mobile mapping and driver assistance systems. This paper proposes a strategy to use road markings as localization landmarks for vision based systems. First step consists in reconstructing a map of road marks. A mobile mapping system equipped with precise georeferencing devices is applied to scan the scene in 3D and to generate an ortho-image of the road surface. A RJMCMC sampler that is coupled with a simulated annealing method is applied to detect occurrences of road marking templates instanced from an extensible database of road mark patterns. The detected objects are reconstructed in 3D using the height information obtained from 3D points. A calibrated camera and a low cost GPS receiver are embedded on a vehicle and used as localization devices. Local bundle adjustment (LBA) is applied to estimate the trajectory of the vehicle. In order to reduce the drift of the trajectory, images are matched with the reconstructed road marks frequently. The matching is initialized by the initial poses that are estimated by LBA and optimized by a MCMC algorithm. The matching provides ground control points that are integrated in the LBA in order to refine the pose parameters. The method is evaluated on a set of images acquired in a real urban area and is compared with a precise ground-truth.