Fusion of Monte Carlo and E2E localizations via importance sampling

Monte Carlo localization (MCL) is a well known model-based localization method and is widely used in many applications. MCL can be easily implemented and its performance is quite stable. However, it cannot cope with unanticipated errors such as wheel slippage because such errors cannot be modeled. On the other hand, end-to-end (E2E) localization that is a learning-based method can cope with unmodelable things if such situations are included in a learning dataset. However, E2E estimate is typically unstable. To realize more stable localization, these advantages must be leveraged without influence of the disadvantages.

In this work, I proposed a hybrid localization method that can simultaneously leverage both the advantages of MCL and E2E while mitigating their disadvantages. To fuse them, I used importance sampling (IS). In the proposal, the learning -based localization method is used to sample particles which are used as candidate of an ego-vehicle in MCL. I used a convolutional network network (CNN) as a learning-based method and Monte Carlo dropout (MCD) is applied to the CNN. MCD enables to regard the CNN as a probabilistic distribution and the particles sampled from the distribution can be fused to MCL via IS without loss of generality.

Fusion process

The gray part of the bottom figure illustrates the fusion process. In MCL, the particles are sampled using the motion model-based proposal distribution referred to predictive distribution. Then, likelihood of the particles is calculated using the measurement model. In the E2E-based sampling, the particles are sampled from the CNN and their likelihood is calculated using the measurement and predictive distributions. The different likelihood calculation manners are derived from IS. Finally, two types of the particles are concatenated and the ego-vehicle pose is estimated. There are no thresholds to switch the use of MCL and E2E; however, this fusion realizes the effective fusion.

Examples of effective fusion

The blue and yellow arrows of the bottom figures show the model- and learning-based particles. In (a), the learning-based particles are sampled at a wrong area; however, the localization has succeeded because the model-based particles are correctly converged. In (b), the model-based particles are not converged, i.e., an initial ego-vehicle pose is not given; however, the learning-based particles are correctly converged. (c) shows a case where the robot moves little bit from (b). The model-based particles are immediately converged by fusing the E2E localization. The proposed localization method can smoothly estimate the robot pose, similar to the model-based method, and quickly re-localize it from the failures, similar to the learning-based method.

Publication

The hybrid localization method is presented in [1].

[1] Naoki Akai, Takatsugu Hirayama, and Hiroshi Murase. "Hybrid localization using model- and learning-based methods: Fusion of Monte Carlo and E2E localizations via importance sampling," In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pp. 6469-6475, 2020. (ResearchGate)