Journal Name:
- International Journal of Science and Engineering Investigations
Author Name |
---|
Abstract (2. Language):
In this paper we introduce a method to detect
abnormal behavior in crowd scenes using spatial-temporal
words, in which Spatial Temporal Interest Points (STIPs) are
extracted as crowd behavior movement features. For this
purpose, three methods for STIPs extraction are compared and
analyzed, which are Harris corner, Gabor wavelet and Hessian
Matrix. Then we select Hessian matrix which can get scale-invariant STIPs. Gradient histogram, optical flow histogram
and spatio-temporal Haar feature are used to build descriptors
for STIPs. In normal behavior modeling Bag-of-words strategy
is used, the keywords of which are produced by GMM based
on EM estimation. Then training samples are divided into
several clips which are described in probability vectors using
keywords, we combine all vectors as a normal behavior
codebook. In recognition phase, we calculate the similarity
distance between the coding vector of the test samples and the
codebook, the abnormal behavior can be detected when the
minimal distance exceeds a threshold. We verify the algorithm
in UMN and UCF datasets, which shows that the proposed
algorithm has effective identification for crowd abnormal
behavior, and it has good robustness against scale variant and
illumination changing
Bookmark/Search this post with
- 3
1-6