Performance metrics and velocity influence for point cloud registration in autonomous vehicles

Abstract: Autonomous vehicles are currently under study and one of the critical parts is the localization of the vehicle in the environment. Different localization methods have been studied over the years, such as the GPS sensor, commonly fused with other sensors such as the IMU. However, situations where the vehicle crosses a tunnel, a bridge, or there is simply traffic congestion, can cause the vehicle to get lost. Therefore, other methods such as point cloud registration have been used, where two point clouds are aligned, thus finding the pose of the vehicle on a precomputed map. Point cloud alignment, although a useful and functional method, is not free from errors that can lead to vehicle mislocalization. The intention of this work is to develop and compare different metrics capable of measuring in real time the performance of the point cloud alignment algorithm used, in this case Normal Distribution Transform (NDT). Therefore, it is important first of all to know if the position obtained meets the minimum requirements defined, just by knowing the input and output parameters of the algorithm. In addition to classifying the positioning as good or bad, the objective is to have a quality parameter that allows estimating the error committed in a complex environment where the uncertainty is very high. In addition, the influence of vehicle speed on the error made by the point cloud alignment algorithm will also be studied to determine whether there is any significant correlation between them. For this purpose, four different metrics have been studied, two of them being new contributions to this algorithm, called Error Propagation and CorAl, while the ones called Hessian and Score are obtained from the alignment algorithm itself. Data used was previously recorded and corrected, therefore obtaining ground truth data. Once the metrics were implemented, all of them were subjected to the same experiments, thus obtaining for each instant a quality measure that allowed a fair comparison to be made. These experiments were carried out on two different routes, being simulated 5 times each. In addition, from these simulations the speed was recorded, allowing the influence study to be carried out. The results show that the best performing metrics in terms of classification and estimation were the Error Propagation and the Hessian, while being impossible to determine a threshold value for the case of CorAl. Furthermore, they show that despite being functional, the error estimation is still far from perfect. It has also been shown that the error estimation of the lateral axis of the vehicle is more complex than in the case of the longitudinal axis. Finally, a strong and positive relationship between the vehicle speed and the error made by the alignment algorithm has been found.