高级检索

基于单目RGB数据的三维模板物体跟踪算法综述

A Survey of 3D Template Object Tracking Algorithms Based on Monocular RGB Data

  • 摘要: 介绍了三维物体跟踪技术及其最新研究进展, 包括理论基础与评估指标、基于视觉特征的传统位姿求解和基于学习的位姿估计. 分析了传统算法中各种特征约束、统计分割模型、能量函数以及加速策略. 其中, 区域及边缘约束成为跟踪的主流方向, 多种特征可以融合产生更强的约束. 统计分割模型由全局模型逐步发展为局部模型. 非线性最小二乘形式的能量函数易于优化而被广泛采用. 预计算、预渲染的加速策略使得跟踪的实时性极大提高. 基于学习的算法现阶段在精度和速度上不及传统算法, 但展现了更好的特征提取能力和处理更复杂场景的潜力. 梳理了多种算法的跟踪精度及取得高精度的原因. 总结了三维物体跟踪在复杂情境下跟踪精度下降和失败的问题, 及其潜在的解决方案与发展方向. 三维物体跟踪正朝向多特征融合、预计算、多任务的方向发展.

     

    Abstract: Introduced the 3D object tracking technology and its latest research progress, including the theoretical foundation and evaluation metrics, traditional pose estimation based on visual features, and learning-based pose estimation. Analyzed various feature constraints, statistical segmentation models, energy functions, and acceleration strategies in traditional algorithms. Region and edge constraints become the mainstream, and fused features can produce stronger constraints. The statistical segmentation model has gradually developed from a global model to local models. The energy function in the form of nonlinear least squares is easy to optimize and is widely used. Pre-computation and pre-rendering acceleration strategies greatly improve the real-time performance. Learning-based algorithms are not as accurate and fast as traditional algorithms, but show greater potential of feature extraction and handling more complex scenes. Summarized tracking accuracy of various algorithms and the reasons for high accuracy. Reviewed the problems of tracking accuracy decline and failure in complex situations, as well as potential solutions and development directions. 3D object tracking is moving towards the direction of multi-feature fusion, pre-computation, and multitasking.

     

/

返回文章
返回