Estimate reliability of an ego-vehicle localization result

Ego-vehicle localization is an essential technique for autonomous navigation. Almost all recent autonomous navigation systems contain the localization module and assume that the localization successfully works always. This assumption is not suitable for safe autonomous navigation; however, it is difficult to detect localization failures because typical localization methods, i.e., optimization- and probability-based methods, do not explicitly estimate criteria regarding the failures. In other words, reliability against the localization result is not estimated in the typical localization framework. Therefore, I propose a novel localization framework that simultaneously estimates the reliability.

Graphical models

The bottom figure shows typical (left) and proposed (right) graphical models, where x is an ego-vehicle pose, z is sensor measurement, u is control input, m is a map. The graphical model represents relationship between probabilistic variables. The typical model only estimates a posterior over the pose conditioned on z, u, and m. Although the posterior describes uncertainty over the pose, it does not describe reliability ("uncertainty" and "reliability" are similar; however, the meanings are completely different!). Hence, the typical model cannot estimate the reliability.

In the proposed graphical model, two variables, decision, d, and localization state, s, are newly introduced. The decision is a probability which indicates whether the localization has failed or not and is output by a classifier which receives the pose, sensor measurement, and map. For example, this classifier is implemented using a deep neural network. However, it is well known that such classifiers cannot output perfect decision. Hence, uncertainty of the decision must be treated. Therefore, the proposed graphical model treats the decision as an observable variable and it depends on the localization state. Where the localization state is a binary variables and indicates "success" and "failure" states. Because the decision depends on the localization state, the localization state is estimated based on the decision. The proposed graphical model estimate a joint posterior over the pose and localization state. Because the probability over the localization state is equal to the reliability, the proposed model can estimate the reliability.

Immediate detection of unreliable localization results

The bottom figure shows an example of performance of the proposed method in the 2D LiDAR-based localization problem. In (a), mismatches between the sensor measurement (red) and map (black) can be seen. The proposed method estimated the reliability less than 95 % and categorized that the localization result is unreliable. In (b), recovering process from the failure, i.e., expand the particle distributions (blue) from the current estimate, was performed. In (c), the localization has succeeded and the proposed method distinguished that the result is reliable. This flow was automatically performed, namely, the robot can recognize whether own localization result is reliable or not by itself.

Publications

The first concept of the proposed localization framework is presented in [1]. I implemented the framework using a convolutional neural network (CNN) and Rao-Blackwellized particle filter (RBPF). The CNN is used as a classifier and it outputs the decision, d, and the RBPF is used to estimate the joint posterior. However, the work presented in [1] has a problem in computational complexity, in particular deep learning-based decision classifier. In [2], I presented an extended implementation and it achieves real time joint posterior estimate. Additionally, I show that the proposed method can be applied to an autonomous cars.

[1] Naoki Akai, Luis Yoichi Morales, and Hiroshi Murase. "Simultaneous pose and reliability estimation using convolutional neural network and Rao-Blackwellized particle filter," Advanced Robotics, vol. 32, no. 17, pp. 930-944, 2018. (ResearchGate)

[2] Naoki Akai, Luis Yoichi Morales, and Hiroshi Murase. "Reliability estimation of vehicle localization result," In Proceedings of the IEEE Intelligent Vehicles Symposium, pp. 740-747, 2018. (ResearchGate)