Proposta de algoritmo de localização por meio de odometria LiDAR

Abstract

In recent decades, there has been a growing interest in the development of autonomous vehicles for various tasks, including domestic, military, and space exploration. The roboticization of complex activities involving navigation in human intervention-free environments implies enabling the agent to acquire models of the scenario and estimate its location in this scene. Thus, it is understood that one of the fundamental needs of an autonomous vehicle is to identify, over time, its location and possess a valid interpretation of the environment in which it is deployed. This technique is known as SLAM (Simultaneous Location and Mapping). In the context of computer vision and mobile robotics, the Iterative Closest Point (ICP) algorithm is widely used in localization and mapping applications. Essentially, this technique is a solution for processing point cloud data, typically obtained from laser sensors, radars, or RGB-D cameras, and can be represented by Cartesian coordinates (3D or 2D points). This work proposes the implementation of the ICP algorithm to process data from a laser scanning sensor (LiDAR; Light Detection and Ranging). It is worth noting that this approach can be used for both localization and mapping; however, the focus of this dissertation is on the localization module. The RMSE (Root Mean Square Error) and ATE (Absolute Trajectory Error) metrics are used to assess the system’s convergence and the error of location parameter estimates, respectively. To validate the trajectory, the ArUco localization module from the OpenCV graphics processing library was used, while the ICP algorithm was developed in C++, aiming at processing on embedded systems. Finally, to assess the possibility of embedding the algorithm’s processing, runtime considerations are evaluated against engineering parameters for a localization application. Although the implemented solution achieves good alignment and an acceptable estimate of the angular component of the location (average error of 6°), translational estimates exhibit a significant error, making the trajectory generated by the approach non-representative of the agent’s actual movement. However, these errors are consistent with the results of a simple implementation approach of the ICP algorithm, which lacks probabilistic filters to correct translational estimates. In addition, it was possible to migrate the system to embedded processing, but the runtime does not satisfy the necessary requirements for a localization system.