Localization failure detection and misalignment recognition

Detection of localization failures is difficult because of the likelihood calculation process in the localization.To solve the localization problem, we need to assume that the sensor measurements are independent to each other. This assumption enables us to decompose the measurement model. Consequently, the measurement model can be calculated in real time. if not, it is infeasible to calculate the model in real time. However, entire relation of the measurements is ignored because the sensor measurements is assumed to be independent. As a result, recognizing misalignment is difficult when partial sensor measurement overlaps with mapped objects. In this work, I proposed a novel method to recognize misalignment and it enables to consider the entire relation.

The right hand figures show recognition results. The top shows a success localization case and the sensor measurements are exactly matched with the map. The red points are points categorized to aligned points. The bottom shows a failure localization case and slight mismatches can be seen. The blue points are points categorized to misaligned points. The proposed method enables to exactly recognize the misalignment while considering the entire relation.

Graphical model

The right figure is the graphical model of Markov random field with fully connected latent variables. In the proposed method, the gray nodes (observable variables), e, are residual errors and white nodes (latent variables), z, are classes of the measurements, i.e., aligned, misaligned, and unknown. Because the latent variables are fully connected to others, the entire relation of the measurements can be considered.