RESUMEN
Monocular depth estimation is a task aimed at predicting pixel-level distances from a single RGB image. This task holds significance in various applications including autonomous driving and robotics. In particular, the recognition of surrounding environments is important to avoid collisions during autonomous parking. Fisheye cameras are adequate to acquire visual information from a wide field of view, reducing blind spots and preventing potential collisions. While there have been increasing demands for fisheye cameras in visual-recognition systems, existing research on depth estimation has primarily focused on pinhole camera images. Moreover, depth estimation from fisheye images poses additional challenges due to strong distortion and the lack of public datasets. In this work, we propose a novel underground parking lot dataset called JBNU-Depth360, which consists of fisheye camera images and their corresponding LiDAR projections. Our proposed dataset was composed of 4221 pairs of fisheye images and their corresponding LiDAR point clouds, which were obtained from six driving sequences. Furthermore, we employed a knowledge-distillation technique to improve the performance of the state-of-the-art depth-estimation models. The teacher-student learning framework allows the neural network to leverage the information in dense depth predictions and sparse LiDAR projections. Experiments were conducted on the KITTI-360 and JBNU-Depth360 datasets for analyzing the performance of existing depth-estimation models on fisheye camera images. By utilizing the self-distillation technique, the AbsRel and SILog error metrics were reduced by 1.81% and 1.55% on the JBNU-Depth360 dataset. The experimental results demonstrated that the self-distillation technique is beneficial to improve the performance of depth-estimation models.
RESUMEN
This article first introduces a sampled-data state estimator design method for continuous-time long short-term memory (LSTM) neural networks with irregularly sampled output. To this end, the structure of the LSTM is addressed to obtain its dynamic equation. As a result, the LSTM neural network is modeled as a continuous-time linear parameter-varying system that is dependent on the gate units. For this system, the sampled-data Luenberger-and Arcak-type state estimator design methods are presented in terms of linear matrix inequalities (LMIs) by using the properties of the gate units. Lastly, the proposed method not only provides a numerical example for analyzing absolute stability but also demonstrates it in practice by applying a pre-trained behavior generation model of a robot manipulator.
RESUMEN
Synchronization between neural networks (NNs) has been intensively investigated to analyze stability, convergence properties, neuronal behaviors and response to various inputs. However, synchronization techniques of NNs with gated recurrent units (GRUs) have not been provided until now due to their complicated nonlinearity. In this paper, we address the sampled-data synchronization problems of GRUs for the first time, and propose controller design methods using discretely sampled control inputs to synchronize master and slave GRUs. The master and slave GRUs are mathematically modeled as a linear parameter varying (LPV) system in which the parameter of the slave GRUs is constructed independently of the master GRUs. This distinctive modeling feature provides flexibility to extend the existing master and slave NNs into a more general structure. Indeed, the sampled-data synchronization can be achieved by formulating the design condition in terms of linear matrix inequalities (LMIs). The novel sampled-data synchronization criteria are devised by combining the H∞ controller design with the looped-functional approach. The synthesized synchronization controllers guarantee not only asymptotic stability of the synchronization error system with aperiodic sampling, but also provides a satisfactory H∞ control performance. Moreover, the communication efficiency is improved by using the proposed method in which the sampled-data synchronization controller is combined with the event-triggered mechanism. Finally, the numerical example validates the proposed theoretical contributions via simulation results.