• simulator
• lidar
• railway
• pose estimation
• odometry
• unreal engine

Nowadays train localization systems are based on infrastructure sensors as transponders (balises) placed between the rails of a railway to capture the position of the train. Optical encoders mounted on the wheels are the only onboard sensor used to measure the velocity of the train and estimate its current position. Any sensor has a specific impairment: wheel sensors suffer from slip and slide effects that become worst with the wear of the wheels; balises are exposed to weather condition and sporadic vandalism that lead to a high maintenance and deployment cost. Thus, balises are not deployed with high density along the railway, resulting in an unsatisfying accuracy in tracing the train position. On the other hand, the position estimated by integrating the velocity measured by the wheel sensors is subject to drift over time, in proportion to the travelled distance and speed. As a consequence, such sensors cannot not guarantee a precise localization of the train between two consecutive balises. Indeed, knowing the exact position of the train would drastically increases the traffic efficiency and the safety of the railway environment.
This thesis has been carried out at the Real-Time Systems Laboratory (ReTiS Lab) of the Scuola Superiore Sant'Anna in collaboration with Hitachi Rail Italy with the objective of designing a novel train localization system. The system must track in real-time the position of the train and must give precise information of which track the train is moving on. Usually, navigation systems are based on Global Navigation Satellite System (GNSS) and Inertial Measurements Unit (IMU) sensors; the GNSS gives absolute position information, while, at a higher frequency either the IMU or Inertial Navigation System (INS) trace the position between two measurements of the GNSS. The railway environment presents several obstacles to satellite signals as narrow valleys, buildings and trees that can lead to signal loss or multi-path effects. Moreover, the INS accumulates drift along the route that prevents a precise localization in GNSS denied environments. Therefore, a multi-sensor system is a promising solution to build a navigation system that can work with high accuracy also when the GNSS signal not available. A Kalman filter can be used to refine the information given by different sensors, such that the accumulated drift of the dead reckoning sensors can be annealed until the next GNSS fix. One of the sensor that is gaining increasing attention in the this research field is the Light Detection and Ranging (LiDAR). It emits light rays that bounce against objects and that are backscattered to the LiDAR receiver that is able to compute the object range relying on the travelling time of the ray. There are several works in the literature that propose various methods to compute the ego-motion of the LiDAR from the captured point clouds; they can be divided in three different classes: registration based algorithms, feature based methods and neural network approaches. Most of them are tested on the KITTI Vision Benchmark Suite that is a dataset composed of different sensor measurements of road trips for automotive purposes, giving a comparison tool for all odometry algorithms based on different sensor variants. This dataset offers real data captured with RGB cameras, stereo cameras, GPS, IMU and LiDAR, giving the chance to test and validate newly proposed algorithms utilizing the GPS/INS as ground truth.
The purpose of this thesis is to investigate and review various approaches to LiDAR odometry presented in the literature. Two distinct LiDAR odometry algorithms have been analyzed and implemented. To evaluate their accuracy, both algorithms have been tested against real sensory data, taken from the KITTI dataset, and synthetic data generated by a train simulator developed under Unreal Engine 4, extended to simulate LiDAR frames. Finally, the results of both the algorithms have been compared.
The thesis is organized as follows. In the first Chapter the functionalities of actual train navigation systems are discussed along with their limitations, then the LiDAR sensor working principles are introduced. Finally, the state-of-the-art LiDAR ego-motion algorithms are presented, as well as the technologies exploited for the implementations. The second Chapter dives into the state-of-the-art algorithms, describing in details their workflows and the main ideas beyond them. In the third Chapter, the implementation details of the chosen algorithms are introduced and described. The fourth Chapter is dedicated to the description of the Unreal Engine environment and to the implementation of the train simulator. In the fifth Chapter, the results and comparisons on both KITTI and simulated dataset are shown and discussed. Finally, Chapter \ref{conclusion} states the conclusions and future development.