一种检测视频中运动目标的新方法

现有的用于视频运动目标检测的鲁棒主成分分析方法通常将背景矩阵的秩函数松弛为核范数，导致求解低秩矩阵的奇异值收缩算子法的阈值恒定，从而背景恢复精度不高。为此提出由加权核范数和结构稀疏范数组成的新的损失函数并用交替方向乘子法进行优化。采用加权核范数作为矩阵的低秩约束，使得压缩阈值与相应奇异值的大小呈单调递减关系，从而大奇异值得以较小幅度压缩。使用结构稀疏范数作为前景稀疏约束，有效利用了前景运动目标的空间区域连续性的先验知识。实验结果表明，该方法在动态背景、阴影等复杂场景下均能取得较其他鲁棒主成分分析方法更好的效果。

A New Method for Detecting Moving Objects in Video

In the method of robust principal component analysis to detection of moving objects, rank function of background matrix is often substituted by the nuclear norm. As a result, the threshold of shrinkage operator of singular values to seek low rank matrix is invariant and the precision of background recovery is relatively low. A new loss function composed of weighted nuclear norm and structured sparsity norm is proposed and optimized by alternating direction multiplier method. In the function, the weighted nuclear norm is adopted as a low rank constraint. It causes that the shrinkage threshold is monotonically decreasing with the corresponding singular value and thus these larger ones are shrunk in a small margin. The structured sparsity norm is used as a sparsity constraint of foreground matrix and it effectively utilizes prior knowledge of the regional continuity of foreground moving objects. The experimental results show that the proposed method can obtain better performances than other robust principal component analysis methods in dynamic background, shadow and other complex scenes.