AI-Driven Sensor Calibration Methods for Autonomous Vehicles

Abstract

With the increasing availability of shared maps and challenges of dynamic scenarios, self-supervised lessons for AVs are becoming more feasible for training data folder and predictive-distributional benchmarks. Sensor fusion has an important role in predictive performance and is directly applicable to dynamic scenarios without the map, for example. Moreover, there are several potential sensors with online learning capabilities that support good contextual information. Required reference and background information for working in this research field are comprised of visual SLAM techniques as well as their role in the validation of node–graph / LiDAR, GNSS and INS intrinsic and extrinsic calibration as given in. Simulation is a popular testing environment, and it is immensely important for transition to the real world without significant infrastructure costs. None of these research areas have the same problems with predictive evaluation.

Autonomous vehicles (AVs) are equipped with sensors for monitoring their surroundings and determining their own state [1]. Most of the studied sensor types can provide local or extended perception, allowing the AV to quickly respond to objects or events nearby [2]. Predictive tasks for which sensors of the vehicle are suitable rely on a combination of these perceptual abilities, and online learning methods are essential to support this. To address AV predictions on a sufficiently broad scale, metrics suitable for evaluating online learning methods must be identified to ensure that training data are acquired in such a way that generalizing to different environments will improve predictive performance. The best predictive task evaluation and training data acquisition metrics are crucial in response to the AV task, but without appropriate sensor fusion, obtained predictive–forward driving performance will inevitably end up being limited [3].