Scholarlink

<p>In self-driving vehicles, Time-of-Flight (ToF) detectors such as LiDAR and RADAR are commonly used to generate depth data for the nearby surroundings. This paper proposes a Ghost Residual Attention Network Based Incorporation of Planar Constraints for Robust Stereo Depth Estimation in Autonomous Vehicle Applications (PCRSDE-GRAN-AVA). Initially, stereo images are collected from the KITTI 2015 dataset. The proposed PCRSDE-GRAN-AVA method predicts affine transformation parameters at the pixel level using depth data from the aggregated cost volume. This is based on the finding that affine transformations can be used to quantify disparities between pixels belonging to planar regions (scene planes) detected in two rectified stereo images. A propagation term is also introduced, ensuring that the same set of constraints is applied to all pixels within the same image plane. The predicted affine parameters are then multiplied by the corresponding pixel coordinates to compute disparity. This step effectively converts the geometric information encoded in affine transformations into depth estimations. The proposed approach is implemented and its efficiency is evaluated using metrics such as MSE, MAD, standardized cross-correlation, loss function and error rate. The efficiency of the proposed PCRSDE-GRAN-AVA technique is compared with existing methods: Deep learning (DL)-based combinations of planar constraints for robust stereo depth estimation in autonomous vehicle applications (PCRSDE-CNN-AVA), Enhanced deep depth assessment for surroundings with thin visual cues (DE-DispResNet-SVC) and Decrypting Pixel Insights: A Deep Dive into DL Models for Improved Indoor Depth Estimation (DPI-U-Net-IDE), respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Robust stereo depth estimation in autonomous vehicle applications by the integration of planar constraints using ghost residual attention networks

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • M. Angelin Ponrani,
  • P. Ezhilarasi,
  • S. Rajeshkannan

摘要

In self-driving vehicles, Time-of-Flight (ToF) detectors such as LiDAR and RADAR are commonly used to generate depth data for the nearby surroundings. This paper proposes a Ghost Residual Attention Network Based Incorporation of Planar Constraints for Robust Stereo Depth Estimation in Autonomous Vehicle Applications (PCRSDE-GRAN-AVA). Initially, stereo images are collected from the KITTI 2015 dataset. The proposed PCRSDE-GRAN-AVA method predicts affine transformation parameters at the pixel level using depth data from the aggregated cost volume. This is based on the finding that affine transformations can be used to quantify disparities between pixels belonging to planar regions (scene planes) detected in two rectified stereo images. A propagation term is also introduced, ensuring that the same set of constraints is applied to all pixels within the same image plane. The predicted affine parameters are then multiplied by the corresponding pixel coordinates to compute disparity. This step effectively converts the geometric information encoded in affine transformations into depth estimations. The proposed approach is implemented and its efficiency is evaluated using metrics such as MSE, MAD, standardized cross-correlation, loss function and error rate. The efficiency of the proposed PCRSDE-GRAN-AVA technique is compared with existing methods: Deep learning (DL)-based combinations of planar constraints for robust stereo depth estimation in autonomous vehicle applications (PCRSDE-CNN-AVA), Enhanced deep depth assessment for surroundings with thin visual cues (DE-DispResNet-SVC) and Decrypting Pixel Insights: A Deep Dive into DL Models for Improved Indoor Depth Estimation (DPI-U-Net-IDE), respectively.