We present a visual SLAM pipeline that is efficient, robust and accurate. It is applied to the trained parking use case. In this case the SLAM algorithm builds a "trained" map on the first pass, typically driven by the driver. In subsequent passes the algorithm localizes to the trajectory, thus allowing the vehicle to autonomously follow the trained path. A visual SLAM system for autonomous vehicles is an attractive option as it utilizes relatively cheap sensors that are typically already mounted on the vehicle for other tasks. However using a visual SLAM approach has challenges, in this paper we specifically look at the localization task in difficult cases. The system is designed to operate in an uncontrolled environment. Between map generation and localization there may be significant changes, different dynamic objects, missing structure, moved structure or the scene may be visually different due to illumination changes or changing weather conditions. These are the so called hard cases. We present an approach, which runs in real-time, designed to tackle the hard cases. The approach has been evaluated both at the bench and in-car.
Over the years, the problem of simultaneous localization and mapping have been substantially studied. Effective and robust techniques have been developed for mapping and localizing in an unknown environment in real-time. However, the bulk of the work presumes that the environment under observation is composed of static objects. In this study, we propose an approach aimed at localizing and mapping an environment irrespective of the motion of the objects in the scene. A hard threshold based Iterative Closest Point algorithm is used to compute transformations between point clouds that are obtained from dense stereo matching. The dynamic entities along with system noise are identified and isolated in the form of outliers of the data correspondence step. A confidence metric is defined that helps in identifying and transitioning a 3D point from static to dynamic and vice versa. The results are then verified in a 2D domain with the aid of a modified Gaussian Mixture Model based motion estimation. The dynamic objects are segmented in 3D and 2D domains for any possible analysis and decision making. The results demonstrate that the proposed approach effectively eliminates noise and isolates the dynamic objects during the mapping of the environment.