Subspace Clustering on sparse component

Introduction

The results of the application of sparse subspace clustering algorithm on raw video for anomaly detection show that anomalies are well separated from normal objects and the background. However, the background and normal sparse components (pedestrians) are not separated well. Some of the points on pedestrians are clustered to background subspace and vice-verse. This may lead to faulty segmentation of anomalies. A better approach would be to separate background before applying the SSC algorithm.

RPCA along with SSC

We first apply the RPCA algorithm discussed in the Robust PCA section to separate the low rank background from the sparse components. The sparse components would contain the normal objects (pedestrians) and the motion anomalies. We then extract keypoints form the sparse components and track them using KLT tracker to obtain the trajectories. Since the number of non-zero features are less, we use Harris operator for keypoint detector instead of SURF as it provides more number of points to be tracked. SSC algorithm applied to the dictionary formed from these trajectory points.

Limitations

The same limitation of SSC of having known number of subspaces lingers here. Although for raw videos we used self-tuning clustering, it doesn't fit here because it clusters the points into a minimum of 2 subspaces. Since we now have only two options for the number of subspaces n = 1 ( non-anomalous case) and 2 (anomalous case), we are bound to choose n = 2 for the SSC algorithm.

Results

The RPCA along with SSC was applied to the following anomalous videos. It can be seen that in the case of anomaly in the scene, it is well segmented out from the other components. But for the frames without anomaly, the algorithm still segments the space into 2. We will be delving deep into this in our future work.