Researches
日本語での簡単な要約はこちら
I participated in Tsukuba Challenge (TC) from 2011 to 2014 as a main robot operator and completed all of the missions. TC is an annual competition of outdoor autonomous mobile robot navigation. Through TC, I developed many modules. Particularly, I mainly developed a localization using a magnetic sensor and LiDAR and person detection modules.
I had studied magnetic map-based localization when I was a master student. The magnetic map-based localization has advantages; however, it also has disadvantages. One of major disadvantage is that building a large-scale magnetic map is significant time consuming. Therefore, I studied how the magnetic map can be quickly and efficiently built when I was a PhD candidate. I proposed the use of Gaussian process regression and LiDAR-based localization techniques for the magnetic map building and achieved large-scale 2D and 3D magnetic field mapping and localization with the dense magnetic map.
From 2016 to 2017, I was developing automated driving modules, particularly 3D LiDAR-based localization and conducted automated driving demonstrations in Japanese public roads. In this term, I worked with Prof. Eijiro Tajeuchi who is a professional of 3D normal distributions transform (NDT) and I also developed 3D NDT-based localization methods.
Recent almost all automated driving systems contain the localization module and assume that the localization successfully works always. This is because that it is difficult to know whether the localization has failed since typical localization methods do not have such a failure detection scheme. Therefore, I proposed a novel localization method that enables to estimate the reliability of the localization result. The proposed localization method enables us to know whether the localization has failed via the estimated reliability.
I studied the 3D LiDAR-based localization for automated driving until 2017. Prof. Takatsugu Hirayama who is a professional of eye-gaze behavior analysis joined our research group since 2018. We discussed about a new research topic and decided to develop a novel platform that enables us to know 3D eye-gaze of a driver on a 3D map (see the right figure). I analyzed relation between eye-gaze and driving behavior around the occluded areas. Additionally, I tried driving behavior modeling with various types of hidden Markov models.
Recognizing misalignment is significant for detecting localization failures; however, it is difficult to recognize the misalignment because entire relation of the sensor measurement cannot be considered in the localization process. I proposed use of Markov random field with fully connected latent variables that enables to consider the entire relation of the sensor measurement via the full connection. The proposed model can recognize slightly mismatched measurements as the misalignment (the blue points in the right figure).
Model- and learning-based localization methods such as Monte Carlo localization (MCL) and end-to-end (E2E) localization have different advantages and disadvantages. To realize an advanced localization system, utilizing both the advantages is a good way; however, there are no suitable framework to fuse these localization methods. In this work, I proposed a hybrid localization method that effectively fuses E2E localization with MCL. The proposed method simultaneously leverages both the advantages while mitigating their disadvantages.
Semantics can be leveraged to improve localization performance because objects with the same class can be matched. However, perfect object recognition is impossible in real environments. Hence, uncertainty of object recognition must be considered in the localization. In this work, I proposed a probabilistic localization method that can cope with uncertainty of object recognition results. The proposed method can accurately and stably estimate the ego-vehicle pose even though the object recognition accuracy is quite bad, e.g., 20 %.
I proposed a stability analysis method of neural networks (NNs) in classification problems. In this work, I assume that the stable NNs do not output outliers while mapping the same class data into a hidden space since output around outliers is drastically changed, i.e., unstable. The proposed method investigates whether the outliers are included in the same class dataset mapped into the hidden space. The key techniques used in the proposed method are persistent homology and its confidences sets estimation method. Simply say, the proposed method enables to investigate whether the mapped dataset is one lump, i.e., topologically simple.
