Geo-Localization Based on Dynamically Weighted Factor-Graph

TL;DR

The paper tackles geo-localization under sparse, ambiguous feature observability and GNSS biases by introducing a dynamically weighted factor-graph that fuses LiDAR-derived information with a GNSS-based prior error model. It formalizes odometry, prior, data association, and prior-error factors, and couples them with a data-information-driven weight update to adapt residual influence in real time. A novel ground-boundary landmark feature, together with ICP-based data association and OpenStreetMap landmarks, enables robust data association even when detections are imperfect. Experimental results on UA campus circuits and KITTI show improved robustness to ambiguities and detection losses, and competitive performance against end-to-end methods, highlighting practical gains for autonomous navigation in challenging environments.

Abstract

Feature-based geo-localization relies on associating features extracted from aerial imagery with those detected by the vehicle's sensors. This requires that the type of landmarks must be observable from both sources. This lack of variety of feature types generates poor representations that lead to outliers and deviations produced by ambiguities and lack of detections, respectively. To mitigate these drawbacks, in this paper, we present a dynamically weighted factor graph model for the vehicle's trajectory estimation. The weight adjustment in this implementation depends on information quantification in the detections performed using a LiDAR sensor. Also, a prior (GNSS-based) error estimation is included in the model. Then, when the representation becomes ambiguous or sparse, the weights are dynamically adjusted to rely on the corrected prior trajectory, mitigating outliers and deviations in this way. We compare our method against state-of-the-art geo-localization ones in a challenging and ambiguous environment, where we also cause detection losses. We demonstrate mitigation of the mentioned drawbacks where the other methods fail.

Figures (7)

• Figure 1: Proposed factor graph: Each factor represents the difference between the observations and the model predictions. Section \ref{['sec:factor_graph']} explains the formalization of the (weighted) factors to obtain the residuals. The empty circles represent the variables that should be estimated by factor residual minimization. Finally, we denote by squares the elements that produce exteroceptive observations, i.e., GPS satellites and landmarks read from the map.
• Figure 2: Example of $\mathcal{D}_i$ and $\mathcal{L}_i$: a) Front-view from LiDAR information where the red line marks the ground boundaries ($\mathcal{D}_i$). b) Front-view mask extracted from (a). c) Top-view from the aerial image where the red line represents the exact boundaries shown in (a) and (b) ($\mathcal{L}_i$).
• Figure 3: Example of $\Phi^{(a)}_i(s_i)$ and $\Phi^{(b)}_i(s_i)$ application in a sigmoid function. The dotted lines indicate the values for $\lambda^{(a)}$ and $\lambda^{(b)}$.
• Figure 4: Aerial image of the University of Alicante Scientific Park, where the evaluation was performed through circuits in Fig. \ref{['fig:circuits']}.
• Figure 5: The ground truth of the four circuits drove around the UA Scientific Park. The landmarks were obtained from the aerial image in Fig. \ref{['fig:aerial']}.