本論文中，我們將分析視訊影像的異常行為偵測。為了簡化我們的問題，我們將問題視為外圍偵測。而在我們測試的資料中，訓練資料皆為正常行為，而測試資料中則是包含了正常和異常的行為。利用訓練資料建立模型，我們可以定義異常行為在模型中正常機率低於閘值。

基於這個概念，我們使用Kullback–Leibler importance estimation procedure (KLIEP)來計算訓練資料和測試資料的比值做為內圍分數。因為在傳統的做法中分別估測兩者機率密度分布的難度很高，所以我們使用KLIEP透過核函式直接模擬兩者之間的比值，可比避免掉分別估測訓練和測試資料的分布。

而後我們利用訓練資料和一些測試資料來建立模型。根據模型，其他的測試資料也可以得到內圍分數，其中內圍分數代表的就是訓練資料和測試資料的相似程度。透過以內圍為基礎的外圍偵測器和內圍分數，我們可以求出適當的閘值，並結合內圍分數來判斷此局部的內圍分數是不是有過低的情形，藉此決定是否在此局部有無異常發生。

在我們的評估中，我們使用PASCAL方法和ground truth來評估局部比值。經過電腦模擬後，我們發現在UCSD的資料和其他相關文獻比較我們的局部比值有著較高的準確性

In this theme, we consider the analysis of abnormally behavior in surveillance system. To simplify the problem, we formalized it as an outlier detection problem. In our case, all behaviors in training data are normal. By creating a model by training data, we can define abnormalities whose probability is below a certain threshold under this model.

Based on this, we use Kullback–Leibler importance estimation procedure (KLIEP) to compute the ratio of training data and testing data which we used as our inlier score. The KLIEP is a method to estimate the inlier score, not the probability densities themselves. This formulation allows us to avoid non-parametric density estimation, which is known to be a difficult task.

After computing inlier score by KLIEP, we create a model by training data and testing data. According to this model, other testing data can also get a inlier score which can represent the degree of similarity between training data and testing data. Based on the concept of inlier-based outlier detection and normal score, we can determine an appropriate threshold and find which location at testing data is abnormal.

In our evaluation we used PASCAL metric to evaluate our localization rate by ground truth. Through computer simulations, we find that our method has high accuracy of localization rate in UCSD dataset compared with previous works.

[1] R. Mehran, A. Oyama, and M. Shah,"Abnormal crowd behavior detection
using social force model,” proc. of CVPR, pp. 935 - 942, June 2009.

[2] D. Helbing, and P. Molna I, "Social force model for pedestrian dynamics,”
Physical Review E, vol. 51, pp. 4282 - 4286, May 1995.

[3] V. Mahadevan, V. B. W. Li, and N. Vasconcelos, "Anomaly detection in
crowded scenes,” proc. of CVPR, pp. 1975 - 1981, June 2010.

[4] Y. Cong, J. Yuan, and J. Liu, "Sparse reconstruction cost for abnormal
event detection,” proc. of CVPR, pp. 3449 - 3456, June 2011.

[5] A. Chan, and N. Vasconcelos, "Modeling, clustering, and segmenting
video with mixtures of dynamic textures,” IEEE Trans. Pattern Anal.
Mach. Intell., vol. 30, pp. 909 - 926, May 2008.

[6] C. Lu, J. Shi, and J. Jia,"Abnormal event detection at 150 fps in matlab,”
proc. of ICCV, Dec. 2013.

[7] Y. Benezeth, P. Jodoin, V. Saligrama, and C. Rosenberger, "Abnormal
events detection based on spatio-temporal co-occurences,” proc. of CVPR,
pp. 2458 - 2465, June 2009.

[8] V. Saligrama, and Z. Chen, "Video anomaly detection based on local
statistical aggregates,” proc. of CVPR, pp. 2112 - 2119, June 2012.

[9] W. Li, M. V., and V. N.,"Anomaly detection and localization in crowded
scenes,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 36, pp. 18 - 32,
June 2013.

[10] M. Sugiyama, T. Suzuki, S. Nakajima, H. Kashima, P. von B Aijnau,
and M. Kawanabe, "Direct importance estimation for covariate shift
adaptation,” Annals of the Institute of Statistical Mathematics, vol. 60,
pp. 699 - 746, Dec. 2008.

[11] Shohei Hido, Yuta Tsuboi, Hisashi Kashima, and Masashi Sugiyama,
"Inlier-based Outlier Detection via Direct Density Ratio Estimation,”
IEEE Trans. Pattern Anal. Mach. Intell., pp. 223-232, Dec. 2008.

[12] R. M.J. and L. M.D.,"Online dominant and anomalous behavior detection
in videos,” proc. of CVPR, pp. 2611 - 2618, June 2013.

[13] N. Dalal, and B. Triggs, "Histograms of oriented gradients for human
detection,” proc. of CVPR, pp. 886 - 893, Jun, 2005.

[14] B.Horn, and B. Schunck, Determining optical flow. ELSEVIER, 1981.

[15] J. Gao, H. Cheng, and P.-N. Tan, "Semi-supervised outlier detection,”
proc. of ACM symposium on Applied Computing, pp. 635–636, 2006.

[16] J. Gao, H. Chengy, and P.-N. Tan,"A novel framework for incorporating
labeled examples into anomaly detection,” proc. of SIAM International
Coference on Data Mining, pp. 593-597, 2006.

[17] J. Yuan, Z. Liu, and Y. Wu,"Discriminative subvolume search for efficient
action detection,” proc. of CVPR, pp. 2442 - 2449, June 2009.

[18] D. Tran, J. Yang, and D. Forsyth,"Video event detection: from subvolume
localization to spatio-temporal path search,” IEEE Tran. Pattern Anal.
Mach. Intell., vol. 36, pp. 404 - 416, July 2013.

[19] T. Ojala, M. Pietikainen, and D. Harwood, "A Comparative Study of
Texture Measures with Classification Based on Feature Distributions,”
Pattern Recognition, vol. 29, pp. 51-59, 1996.

[20] J. Fehr, and H. Burkhardt,"3D rotation invariant local binary patterns,”
in ICPR, pp. 1-4, Dec. 2006.

[21] J. Yuan, Z. Liu, and Y. Wu,"Discriminative subvolume search for efficient
action detection,” proc. of CVPR, pp. 2442–2449, 2009.

[22] W. Hardle, M. Muller, S. Sperlich, and A. Werwatz, Nonparametric and
Semiparametric Models. Springer Series in Statistics, 2004.

[23] K.-W. Cheng, Y.-T. Chen, and W.-H. Fang, "Abnormal crowd behavior
detection and localization using maximum sub-sequence search,” proc. of
ACM/IEEE international workshop on Analysis and retrieval of tracked
events and motion in imagery stream, pp. 49-58, Oct, 2013.