Lost! Leveraging the Crowd for Probabilistic Visual Self-Localization

Abstract

In this paper we propose an affordable solution to selflocalization, which utilizes visual odometry and road maps as the only inputs. To this end, we present a probabilistic model as well as an efficient approximate inference algorithm, which is able to utilize distributed computation to meet the real-time requirements of autonomous systems. Because of the probabilistic nature of the model we are able to cope with uncertainty due to noisy visual odometry and inherent ambiguities in the map (e.g., in a Manhattan world). By exploiting freely available, community developed maps and visual odometry measurements, we are able to localize a vehicle up to 3m after only a few seconds of driving on maps which contain more than 2,150km of drivable roads.

Robust monocular visual odometry for road vehicles using uncertain perspective projection

“…The ability to correctly estimate big maneuvres is especially important for the integration with offline map data, as maneuvres are generally more reliable clues for map matching than gentle curves and straights. The concept of map matching as a mechanism to eliminate error accumulation has already been proven [2,23].…”

“…1 Ghent University -IPI/iMinds, St-Pietersnieuwstraat 41, B-9000 Ghent, Belgium. 2 Grontmij Belgium, Arenbergstraat 13/1, B-1000 Brussels, Belgium. …”

“…The greater the distance traveled from the last absolute reference point (e.g., the known GPS coordinates of the starting address), the greater the positional error can become. While this may appear to limit the use of visual odometry to a short distances, the drift error can be bounded by combination with additional passive sensors (e.g., a magnetic compass for dead reckoning) or a priori known information (e.g., the local road map) [2,3]. As such, visual odometry is still a prime candidate to supplement satellite navigation even for urban scenarios where signal reception may be unreliable for prolonged distances.…”

Topometric localization on a road network

“…Localization is achieved by matching observed trajectories, constructed from visual odometry [9], [12] or IMU-based odometry [22], with a prior map using various techniques, such as Chamfer matching [12], curve similarity matching [22], or street segment matching [9]. Such approaches, however, are subject to drift in situations where the road geometry offers no constraints (e.g., long, straight road segments), particularly when visual odometry is used.…”

“…Metric localization provides, instead, quantitative estimates of observer position in a map. Simultaneous localization and mapping (SLAM) [17], [18], and visual odometry [9], and some appearance-based localization approaches relying on accurate 3D maps [? ], [15] fall into this category.…”