1

我正在尝试使用移动摄像头跟踪车辆。车辆由OpenCV 中基于 Haar 特征的级联分类器检测。我在这里使用预训练数据集。根据对象跟踪调查（2006 年），跟踪对象分为三类：点或邻域、内核或形状，或者跟踪可以基于对象的轮廓或其允许者。我认为我的方法属于第二类。因此，我首先尝试了 camshift 算法。

然而，在检测阶段，我得到了很多误报，这使得跟踪效率低下。我的问题是如何消除误报？

我正在考虑为每一帧计算所有边界框（包括误报）的直方图。然后，对下一帧执行相同操作并比较直方图以创建置信度。你认为这是个好主意吗？

我也有想法在检测之前应用处理（例如：分割背景/前景减法，阈值等），但这些更适合固定相机而不是移动相机。我需要继续和 Haar 合作一段时间，至少试一试。任何建议都会很棒。

4

1 回答 1

1

This is actually a broad question with a myraid of approaches, but the basic one is that you need a few things to build a solid tracker:

  • a good input (moving camera)
  • a good feature / region descripter (you chose haar wavelet based method)
  • a good comparison metric for discripters (don't know what you are doing here)

generically you need to stabilize your input through an image registration process reference your library of discripters calculate a metric of those discripters against your registered input image

reducing false positives is dependent upon your approach, but typically you group together the false positives and find out where in the algorithm they are originating from the most (many will originate at different places) and then change that part of the algorithm to cull them better.

your options are:

  • register the images if they aren't already (this will remove the false positives from relative motion)
  • examine the algorithm to identify the largest source of false positives and modify it accordingly
  • get better training data to match against (if your training data doesn't match the real world data this will never work)
  • try a different approach (SIFT,SURF, GLO, Fourier Mellin Transform etc.)

that's as specific as I can be given your provided information. I hope it helps.

于 2014-04-24T13:17:39.480 回答