摘要： 大多数车载行人方案是基于特征选择和机器学习的，但大量特征的计算使实时性大幅降低。为将待检测窗口限制在最小的范围内，提出一种改进的基于立体视觉的摄像机角度估计自适应图像采样方法，利用基于类Haar特征和Real AdaBoost学习方法的分类器进行实现，在户外移动平台上对处于复杂动态背景中的行人目标进行检测。结果表明，与其他方法相比，该方法在保证检测效果的同时，计算时间仅为自适应路面拟合方法的13%。

关键词: 驾驶系统, 行人检测, 摄像机角度估计, 类Haar特征, Real AdaBoost训练

Abstract: On-board pedestrian detection needs processing scenarios from a mobile platform, which implies environments change quickly. And the aspects of pedestrian are inconstant, which makes detection difficult. Many approaches based on machine learning use a large number of features which need much computing time. To improve the problem, this paper provides an improved camera pose estimation method for adaptive sparse image sampling, and a classifier based on Haar-like wavelets and Real AdaBoost as learning machine. It compares the proposal with relevant approaches, experimental results show that the method reduces processing time much a lot for the image sampling.

Key words: drive system, pedestrian detection, camera pose estimation, Haar-like feature, Real AdaBoost training

中图分类号: 

• TP18