Guillaume Bresson

Real-Time Monocular SLAM With Low Memory Requirements

The localization of a vehicle in an unknown environment is often solved using simultaneous localization and mapping (SLAM) techniques. Many methods have been developed, each requiring a different amount of landmarks (map size), and thus of memory, to work efficiently. Similarly, the required computational time is quite variable from one approach to another. In this paper, we focus on a monocular SLAM problem and propose a new method called MSLAM, which is based on an extended Kalman filter (EKF). The aim is to provide a solution that has low memory and processing time requirements and that can achieve good localization results while benefiting from EKF advantages (i.e., direct access to the covariance matrix, no conversion required for the measures or the state, etc.). To do so, a minimal Cartesian representation (three parameters for three dimensions) is used. However, linearization errors are likely to happen with such a representation. New methods allowing to avoid or hugely decrease the impact of the linearization failures are presented. The first contribution proposed here computes a proper projection of a 3-D uncertainty in the image plane, allowing to track landmarks during longer periods of time. A corrective factor of the Kalman gain is also introduced. It allows to detect wrong updates and correct them, thus reducing the impact of the linearization on the whole system. Our approach is compared with a classic SLAM implementation over different data sets and conditions to illustrate the efficiency of the proposed contributions. The quality of the map built is tested by using it with another vehicle for localization purposes. Finally, a public data set presenting a long trajectory (1.3 km) is also used in order to compare MSLAM with a state-of-the-art monocular EKF-SLAM algorithm, both in terms of accuracy and computational needs.

Parsimonious real time monocular SLAM

This paper presents a real time monocular EKF SLAM process that uses only Cartesian defined landmarks. This representation is easy to handle, light and consequently fast. However, it is prone to linearization errors which can cause the filter to diverge. Here, we will first clearly identify and explain when those problems take place. Then, a solution, able to reduce or avoid the errors involved by the linearization process, will be proposed. Combined with an EKF, our method uses resources parsimoniously by conserving landmarks for a long period of time without requiring many points to be efficient. Our solution is based on a method to properly compute the projection of a 3D uncertainty into the image frame in order to track landmarks efficiently. The second part of this solution relies on a correction of the Kalman gain that reduces the impact of the update when it is incoherent. This approach was applied to a real data set presenting difficult conditions such as severe distortions, reflections, blur or sunshine to illustrate its robustness.

Simultaneous Localization and Mapping: A Survey of Current Trends in Autonomous Driving

In this paper, we propose a survey of the Simultaneous Localization And Mapping (SLAM) field when considering the recent evolution of autonomous driving. The growing interest regarding self-driving cars has given new directions to localization and mapping techniques. In this survey, we give an overview of the different branches of SLAM before going into the details of specific trends that are of interest when considered with autonomous applications in mind. We first present the limits of classical approaches for autonomous driving and discuss the criteria that are essential for this kind of application. We then review the methods where the identified challenges are tackled. We mostly focus on approaches building and reusing long-term maps in various conditions (weather, season, etc.). We also go through the emerging domain of multivehicle SLAM and its link with self-driving cars. We survey the different paradigms of that field (centralized and distributed) and the existing solutions. Finally, we conclude by giving an overview of the various large-scale experiments that have been carried out until now and discuss the remaining challenges and future orientations.

A cooperative fusion architecture for robust localization: Application to autonomous driving

The localization of a vehicle is a central task of autonomous driving. Most of the time, it is solved by considering a single algorithm with a few sensors. In this paper, we propose a cooperative fusion architecture based on two main algorithms: a laser-based Simultaneous Localization And Mapping (SLAM) process and a lane detection and tracking approach using a single camera. Both algorithms are designed individually as cooperative fusion processes where other sensors (GPS and proprioceptive information) and dedicated maps are integrated to strengthen the advantages of each system. The whole architecture is formalized around key components (ego-vehicle, roadway, obstacle and environment). A final decision layer, that takes into account the state of each algorithm, allows the system to choose the most appropriate ego-vehicle localization mean based on the current road situation and the environmental context.

Monocular urban localization using street view

This paper presents a metric global localization in the urban environment only with a monocular camera and the Google Street View database. We fully leverage the abundant sources from the Street View and benefits from its topo-metric structure to build a coarse-to-fine positioning, namely a topological place recognition process and then a metric pose estimation by local bundle adjustment. Our method is tested on a 3 km urban environment and demonstrates both sub-meter accuracy and robustness to viewpoint changes, illumination and occlusion. To our knowledge, this is the first work that studies the global urban localization simply with a single camera and Street View.

A general consistent decentralized Simultaneous Localization And Mapping solution

In this paper, we propose a new approach to the decentralized Simultaneous Localization And Mapping (SLAM) problem. The goal is to demonstrate the feasibility of decentralized localization using low-density maps built with low-cost sensors. This problem is challenging at different levels. Indeed, each vehicle localization tends to drift over time independently of one another making the global localization of a fleet hard to achieve. To counter this effect, called SLAM inconsistency and which has been stated numerous times in the literature, we introduce a model to represent the natural drift of SLAM algorithms. Its integration inside an Extended Kalman Filter is explained along with simulations validating its use. The second part of this paper presents the fusion architecture designed to solve the different problems arising in a decentralized scheme. It avoids data incest, which is an important source of inconsistency, and integrates the previously mentioned SLAM drift in the estimates produced. This architecture also separates the SLAM classically used for mono-vehicle applications from the high-level decentralized part offering flexibility regarding sensors and algorithms at a low-level. Other aspects, involved by the multi-vehicle settings, are also taken into account (communication losses, latencies, desynchronizations, unknown initial positions of the fleet members and data association). The whole algorithm has been tested in various scenarios with vehicles equipped with a single camera and an odometer. The results, from both simulated and real scenarios, show that our approach can work in real time with very small bandwidth requirements. A fully decentralized SLAM algorithms based on a monocular setting.A general architecture handling data incest and network aspects.The integration of a model for the natural SLAM drift that ensures consistency.Unknown initial positions and the drift are resolved by sharing maps within the fleet.The real time application of our approach to various scenarios with 2 or 3 vehicles.

Making visual SLAM consistent with geo-referenced landmarks

This paper presents a solution to the consistency problem of SLAM algorithms. We propose here to model the drift affecting the estimation process. The divergence is seen as a bias on the vehicle localization. By using such a model, we are able to guarantee the consistency of the localization. We developed a filter taking into account the divergence and allowing to easily integrate any information helping to characterize the current drift. Geo-referenced landmarks are used in order to provide an absolute localization and drastically reduce the impact of the divergence. The filter is designed around an Extended Kalman Filter and is totally separated from the classical SLAM algorithm. Our method can consequently be connected to any existing SLAM process without trouble. A vehicle performing monocular SLAM in real time was used to validate our approach with real data. The results show that the integrity of the filter is preserved during the whole trajectory and that geo-referenced information helps reducing the natural SLAM drift.

Loop Closing in a Drift-Aware Monocular SLAM

Abstract This paper presents a real-time Consistent Monocular EKF-SLAM process. We introduce the notion of bias which allows to model the natural drift of the SLAM process. Thanks to it, the consistency of the filter is guaranteed. By connecting the bias to the different landmarks and to the vehicle pose, the estimates become tightly bound to the SLAM drift. It means that a loop closure, for instance, will naturally estimate the bias and so correct the vehicle pose and landmark positions without any special processing. We developed a dedicated architecture in order to integrate the bias. It uses an Extended Kalman Filter and has the advantage to be totally decorrelated from the classical SLAM process. Thanks to it, any algorithm, with any kind of sensors or methods, can be used instead of the monocular SLAM employed in this paper, as long as it produces landmark estimates and their uncertainty. This approach was validated with a real experiment composed of a long loop.

Improving robustness of monocular urban localization using augmented Street View

With the fast development of Geographic Information Systems, visual global localization has gained a lot of attention due to the low price of a camera and the practical implications. In this paper, we leverage Google Street View and a monocular camera to develop a refined and continuous positioning in urban environments: namely a topological visual place recognition and then a 6 DoF pose estimation by local bundle adjustment. In order to avoid discrete localization problems, augmented Street View data are virtually synthesized to render a smooth and metric localization. We also demonstrate that this approach significantly improves the sub-meter accuracy and the robustness to important viewpoint changes, illumination and occlusion.

Improving SLAM with Drift Integration

Localization without prior knowledge can be a difficult task for a vehicle. An answer to this problematic lies in the Simultaneous Localization And Mapping (SLAM) approach where a map of the surroundings is built while simultaneously being used for localization purposes. However, SLAM algorithms tend to drift over time, making the localization inconsistent. In this paper, we propose to model the drift as a localization bias and to integrate it in a general architecture. The latter allows any feature-based SLAM algorithm to be used while taking advantage of the drift integration. Based on previous works, we extend the bias concept and propose a new architecture which drastically improves the performance of our method, both in terms of computational power and memory required. We validate this framework on real data with different scenarios. We show that taking into account the drift allows us to maintain consistency and improve the localization accuracy with almost no additional cost.