Machine Vision Technology：Lecture10 Object Detection

- [object detection and challenges](#object detection and challenges)
- [Face Detection人脸检测](#Face Detection人脸检测)
- [Sliding Window Face Detection with Viola-Jones 2001](#Sliding Window Face Detection with Viola-Jones 2001)
- [Pedestrian Detection行人检测](#Pedestrian Detection行人检测)

计算机视觉（本科）北京邮电大学鲁鹏

Introduction of object detection

Face Detection

Pedestrian Detection

object detection and challenges

Object Detection Design challenges：

How to efficiently search for likely objects
- Even simple models require searching hundreds of thousands of positions and scales.即使是简单的模型也需要搜索成千上万的位置和尺度。
Feature design and scoring 特征设计和评分
- How should appearance be modeled?
- What features correspond to the object?
How to deal with different viewpoints? 处理不同的视角差异
- Often train different models for a few different viewpoints 经常为几个不同的视角训练不同的模型

Face Detection人脸检测

Challenges of face detection：

Sliding window = tens of thousands of location/scale evaluations
- 一个百万像素的图像有大约 1 0 6 10^6 106 个像素，以及相当数量的候选人脸位置
Faces are rare: 0--10 per image 人脸罕见：每张图片0-10哥人脸
- For computational efficiency, spend as little time as possible on the non-face windows. 为了提高计算效率，在非人脸窗口上花费尽可能少的时间。
- For 1 Mpix, to avoid having a false positive in every image, our false positive rate has to be less than 1 0 − 6 10^{-6} 10−6 1Mpix,避免假阳性在每一个图像,我们的假阳性率必须小于 1 0 − 6 10^{-6} 10−6

Sliding Window Face Detection with Viola-Jones 2001

Viola-Jones使用了机器学习的boosting算法，下面是boosting算法介绍：

A simple algorithm for learning robust classifiers
Provides efficient algorithm for sparse visual feature selection
Easy to implement, not requires external optimization tools

1.找到正确率大于0.5的分类器 h i ( x ) h_i(x) hi(x)

2.把错误分类的权重放大

3.迭代1-2

通过几个分类器组合起来得到最终分类器。
h ( x ) = α 1 h 1 ( x ) + α 2 h 2 ( x ) + α 3 h 3 ( x ) + ⋯ h(x) = \alpha_1 h_1(x) + \alpha_2 h_2(x) + \alpha_3 h_3(x) + \cdots h(x)=α1h1(x)+α2h2(x)+α3h3(x)+⋯

其中 h ( x ) h(x) h(x) 是 Strong classifier 强分类器， h i ( x ) h_i(x) hi(x) 是Weak classifier， x x x 是 Features vector， α i \alpha_i αi 是 Weight。

每个弱分类器：
h j ( x ) = { 1 if f j ( x ) > θ j 0 otherwise h_j(x) = \left\{ \begin{array}{rcl} 1 & & \text{if} \quad {f_j(x) \gt \theta_j} \\ 0 & & \text{otherwise} \\ \end{array} \right. hj(x)={10iffj(x)>θjotherwise

其中 f j ( x ) f_j(x) fj(x) 是 value of rectangle feature， θ j \theta_j θj 是threshold。如下图所示。

所以最终的 strong classfier：
h ( x ) = { 1 ∑ t = 1 T α t h t ( x ) > 1 2 ∑ t = 1 T α t 0 otherwise h(x) = \left\{ \begin{array}{rcl} 1 & & {\sum\limits_{t = 1}^{T} \alpha_t h_t(x) \gt \frac{1}{2} \sum\limits_{t = 1}^{T} \alpha_t } \\ 0 & & \text{otherwise} \\ \end{array} \right. h(x)=⎩ ⎨ ⎧10t=1∑Tαtht(x)>21t=1∑Tαtotherwise

Viola & Jones algorithm ：

A "paradigmatic" method for real‐time object detection 实时目标检测的"范例"方法
Training is slow, but detection is very fast
Key ideas：
- Integral images for fast feature evaluation 用于快速特征评估的积分图像（积分图）。
- Boosting for feature selection
- Attentional cascade for fast rejection of non‐face windows 注意力级联快速拒绝非人脸窗口。也就是对非人脸窗口处理用时更少。

详情看论文，没听太明白。。。就积分图有点像二维联合分布函数的矩形公式。

Pedestrian Detection行人检测

Histograms of oriented gradients for human detection 2005

HoG Feature.