论文解析之运动检测篇 —《Learning Patterns of Activity Using Real-Time Tracking》

Goals：

• Self-calibration: determine the positions of all the cameras relative to one another;
• Construction of rough site models: determine the ground plane and mark occupied areas;
• Detect objects in the site: extract information about all moving objects in the site;
• Classify detected objects: label detected objects by common shape, appearance, or motion;
• Learn from extended observation: what are the common activity patterns;
• Detect unusual events in the site: mark activities that don’t fit common patterns.

Problems：

• Cluttered areas
• Objects overlapping
• Shadows
• Lighting changes
• Moving elements of the scene
• Slow-moving
• Objects introduce or remove

Relate Work

• Maximum interframe difference[^1]
• Adaptive Gaussian statistical model[^2]
• Disparity verification[^3][4]
• Kalman Filter[^5]
• Multiclass statistical model[^6]
• GMM[^7]

Approach

Parameters

• the learning constant $\alpha$
• the proportion of the data $T$

Gaussian Distributions

$$P(X_t)=\sum _{i=1}^K{w}_{i,t}\ast \eta (X_t,{\mu }_{i,t},{\Sigma }_{i,t})$$

• 模型参数：
  $K$ is the number of distributions, $w_{i,t}$ is an estimate of the weight (the portion of the data accounted for by this Gaussian) ,$\eta$ is a Gaussian probability density function.
• 计算简化：
  $K$=3~5; ${\Sigma }_{k,t}={\sigma }_k^{2}I$; EM$\to$K-Means
• 模型匹配：
  阈值：2.5$\sigma$ （不同分布阈值不同）
  If none of the $K$ distributions match the current pixel value, the least probable distribution is replaced with a distribution with the current value as its mean value, an initially high variance, and low prior weight.
• 模型更新
  • 权重更新：$w_{k,t}=(1-\alpha )w_{k,t-1}+\alpha (M_{k,t})$
  • 参数更新：${\mu }_t=(1-\rho ){\mu }_{t-1}+\rho X_t,{\sigma }_t^2=(1-\rho ){\sigma }_{t-1}^2+\rho (X_t-{\mu }_t{)}^T(X_t-{\mu }_t),\rho =\alpha \eta (X_t|{\mu }_k,{\sigma }_k)$
• 背景估计：
  Most supporting evidence and the least variance, listed by $w/\sigma$.
  $$B={\arg \min }_b(\sum _{k=1}^b{w}_k>T)$$

Tracker (frame):

• 连通分支[^8]：Determine the whole moving object, and character their position, size, moments, and other shape information.
• 卡尔曼模型池 + 连通分支池
• 模型依概率匹配到连通域
• 误差较小的匹配用于更新对应模型
• 更新后的模型如果具备足够的适合度，应用到下一帧
• 无法匹配的连通域搜索并建立新的卡尔曼模型
• 使用前两帧不匹配的连通分支建立新模型
• 使用当前帧和前两帧中未匹配的模型假设新模型
• 如果没有模型匹配为空，则进行常数弱化传播
• 去除适应度（预测误差的方差倒数）低于阈值的模型
• 跟踪单个对象的连续序列，发生交互停止跟踪
• 删除可能性最低模型防止组合爆炸