COMP9517 week7 Motion

https://echo360.org.au/lesson/9750cbfe-6500-42f7-b024-45c5cf2a69b0/classroom#sortDirection=desc

https://webcms3.cse.unsw.edu.au/static/uploads/course/COMP9517/20T2/bca52a9037053fc7f77869b5c36edef5830da7920d455dce479c282c9ba12431/COMP9517_20T2W7_Part_1-1_Motion.pdf

Adding the time dimension to the image formation

1. Detecting Motion 

  1). Change Detection 

    image subtraction:  两图相减,对于每个pixel的值, 大于threshold为1,小于为0

    

  2).  Sparse motion estimation Using template matching to estimate local displacements

    (1) Motion Vector : 向量代表点的移动

      

       (2) 检测目标点, 寻找对应点在t+Δt ; 假设点移动的不远。

        

    (3) 检测目标点的方法 filters, operators

       计算水平,垂直,斜线方向的intensity variance, 如果最小值>threshold,则是interest point

      

     (4)Search corresponding points

      template searching:

        1. 对于 interest point,在其周围划定一个neighbour区域,基于一些prior knowledge,在下一张图限定search region,try to find a similar pattern

        2. Similarity Measures : 对两个template,计算相似度:

          1) 计算对应点乘积之和,求最大 ; 计算对应点绝对值之差,求最小; 计算对应点之差的平方,求最小

          2)Mutual information,求最大:   intensity distribution histogram 计算 概率

      

       

  3).  Dense motion estimation Using optical flow to compute a dense motion vector field

    (1)假设:time interval之间,物体与camera之间不产生巨大位移; t时刻的neighbourhood能在t+Δt时刻找到

    (2)联立 泰勒展开与OF不等式

      

       

     

 

2. Tracking 

原文地址:https://www.cnblogs.com/ChevisZhang/p/13340392.html