Towards Accurate Active Camera Localization

For robots,autonomous vehicles,and other intelligent agents,obtaining a clear pose in the scene is a fundamental requirement for completing many tasks.And for augmented reality and virtual reality tasks,accurate localization helps to anchor virtual objects more accurately to the real environment,achieving a fusion of the real and virtual worlds.Most previous camera localization work has focused on passive camera localization,which directly predicts the camera pose using the visual features and geometric structure of the current visual image.However,in specific cases such as sparse texture or lighting changes,this type of prediction is difficult to accomplish.Therefore,some work attempted to tackle the active camera localization problem which allows cameras to actively move the camera to obtain better localization results.However,these methods rely on discrete belief maps and face problems such as low localization accuracy and poor efficiency.This article focuses on the above issues faced by current active localization algorithms.To address the accuracy issue,this article proposes to optimize the camera pose by the connection between local coordinates and world coordinates in the continuous pose space,instead of using the Markov localization in the discrete space as the previous active camera localization algorithms.To address the efficiency issue,this article calculates the localization accuracy at different scales,constructs scene uncertainty maps of different scales,and uses them to assist camera movement.At the same time,considering that previous active camera localization algorithms lack a good method for calculating the current camera uncertainty,making it difficult for the algorithm to choose a specific stop time,this article proposes a new method for calculating camera uncertainty,enabling the camera to choose the stop time more intelligently and further improve the efficiency.This article validates the effectiveness of the proposed algorithm on high-difficulty camera localization problems in both synthetic scenes and indoor scenes obtained through scanning and reconstruction.In the experiments,this article not only compared with the best existing active localization algorithms but also further compared with the heuristic baselines based on passive camera localization algorithms,and designed ablation experiments to demonstrate the effectiveness of each module of the proposed algorithm.This article achieved a localization success rate of 83.05%and 82.40%in synthetic and scanned scenes,respectively,under the success conditions of 5cm and 5°,surpassing all other comparison methods.At the same time,compared with other active localization algorithms,the proposed algorithm can terminate the movement in advance,with higher efficiency.