Search

This research proposes a method for accurate all-round 3D reconstruction of an indoor environment in one-shot using a system of trinocular 360-degree cameras. Binocular 360-degree stereo is unable to reconstruct in all directions due to lack of disparity along epipolar directions. Thus, a third camera along a perpendicular epipolar direction is introduced to cover for this, making the system trinocular. However, previous works with trinocular stereo did not adequately take into account the uncertainty of disparity estimation and geometric constraints around 3D reconstruction. Therefore, we propose a geometric optimization scheme considering disparity estimation uncertainty and show that this results in both higher accuracy and lesser outliers along epipolar directions, in both simulated and real environments. [ABSTRACT FROM AUTHOR]

Semantic information has started to be used in localization methods to introduce a non-geometric distinction in the environment. However, efficient ways to integrate this information remain a question. We propose an approach for fusing data from different object classes by analyzing the posterior for each object class to improve robustness and accuracy for self-localization. Our system uses the bearing angle to the objects' center and objects' class names as sensor model input to localize the user on a 2D annotated map consisting of objects' class names and center coordinates. Sensor model input is obtained by an object detector on equirectangular images of a 360° field of view camera. As object detection performance varies based on location and object class, different object classes generate different likelihoods. We account for this by using appropriate weights generated by a Gaussian process model trained by using our posterior analysis. Our approach follows a systematic way to fuse data from different object classes and use them as a likelihood function of a Monte Carlo localization (MCL) algorithm. [ABSTRACT FROM AUTHOR]

This paper proposes a method to improve the accuracy of 3D measurement of a stereo camera by marking a measured object using a line laser. Stereo cameras are commonly used for 3D measurement, but the accuracy of 3D measurement is affected by the amount of texture. Therefore, a new measurement system combining a stereo camera and a line laser is developed. The accuracy of 3D measurement with a stereo camera is improved by using a line laser to mark arbitrary points on the measured object and measuring the marked points, regardless of the amount of texture on the measured object. Because the laser is only used to mark points on the measurement target, calibration is not required with the stereo camera. Experimental evaluation showed that our proposed method can obtain millimeters. [ABSTRACT FROM AUTHOR]

In this paper, we propose a novel algorithm for estimating road surface shapes and object heights using a fisheye stereo camera. Environmental recognition is an important task for advanced driver-assistance systems. However, previous studies have only achieved narrow measurement ranges owing to sensor restrictions. Moreover, the previous approaches cannot be used in environments where the slope changes because they assume inflexible constraints on the road surfaces. We use a fisheye stereo camera capable of measuring wide and dense 3D information and design a novel algorithm by focusing on the degree of division in a disparity image to overcome these defects. Experiments show that our method can detect an object in various environments, including those with inclined road surfaces. [ABSTRACT FROM AUTHOR]

In this paper, we propose a self-supervised optical flow-based approach to learn the rotation of an arbitrarily moving spherical camera. Nowadays, deep learning has enabled efficient learning of camera rotation efficiently. However, most approaches are fully supervised and require large datasets with ground-truth labels of the rotation, and these labels are difficult to acquire. We attempt to solve this problem by using a derotation operation of the spherical optical flow on a unit sphere. This operation decouples the camera rotation from the mixture of translational and rotational components, removing the effect of 3D information for rotation estimation. Therefore, we integrate a derotation layer into a convolutional neural network for regressing the camera rotation. This layer can be adopted for only spherical cameras, which can capture all-round information, and thus enables the network to be learned the camera rotation without using labeled training datasets. We experimentally demonstrate that our approach achieves the comparable performance for the rotation estimation to that of a fully supervised approach and that it outperforms a previously proposed approach. Moreover, transfer learning is conducted in new environments to confirm the benefit of the self-supervised learning. [ABSTRACT FROM AUTHOR]