Unsual Behavior Analysis and Its Application to Surveillance Systems Yung-Tai Hsu( 許詠泰 ) Jun-Wei Hsieh( 謝君偉 ) Hong-Yuan Mark Liao( 廖弘源 )

Introduction Deformable Triangulations Skeleton-based Posture Recognition Posture Recognition Using the Centroid Context Experiment Results

Deformable Triangulations P is a posture extracted in binary form by image subtraction(fig.1) B is the set of boundary points along the contour of P(fig.2) Extract some high curvature points from B(fig.3) α(p) is the angle of a point p in B. It can be determined by two specified points p + and p -.(fig.4) fig.1 fig.2 fig.3 fig.4

Deformable Triangulations D min = |B| / 30, D max = |B| / 20 If α is larger than a threshold T α (here we set it at 150), p is selected as a control point. If two candidates, p 1 and p 2 are close to each other, i.e., ||p 1 – p 2 ||<d min, the candidate with smaller α angle is chosen as a control point.

Deformable Triangulations Vi Vj Vk VaVb

Triangulation-based Skeleton Extraction P is decomposed into a set of triangle meshes Ω p Ω p ={T i } i=0,1,2,…,N TP -1 Each triangle mesh T i in Ω p has a centroid C T i H is defined as the head of P and it is the highest node among all the nodes. All the leaf nodes L i correspond to different limbs of P The branching nodes B i are the key points used to decompose P into different body parts, such as the hands, feet, or torso.

Posture Recognition Using a Skeleton Assume S P and S D are two skeletal images extracted from a posture P and D. Assume DT S P is the distance map of S P. The value of a pixel r in DT S P is its shortest distance to all foreground pixels in S P. d(r, q) is the Euclidian distance between r and q. |DT S P | represents the image size of DT S P. S P and S D must be normalized to a unit size and their centers must be set to the origins of DT S P and DT S D.

Posture Recognition Using a Skeleton a)Shows the original posture. b)It is the result of skeleton extraction. c)Shows the resultant distance map based on (b)

Centroid Context-based Description of Postures Assume all postures are normalized to a unit size. We project a sample onto a log-polar coordinate and label each mesh. Use m to represent the number of shells used to quantize the radial axis and use n to represent the number of sectors that we would like to quantize each shell. The total number of bins used to construct the centroid context is m×n. For each centroid r of a triangle mesh of a posture, we construct a vector histogram h r. h r (k) is the number of triangle mesh centroids in the kth bin by considering r as the origin bin k is the kth bin of the log-polar coordinate.

Centroid Context-based Description of Postures Given two histograms h r i (k) and h r j (k), the distance between them can be measured by a normalized intersection:

Centroid Context-based Description of Postures |V P | is the number of elemetns in V P.

Centroid Context-based Description of Postures Give two postures P and Q, the distance between their centroid contexts is measured by: Where w and w are the area ratios of the ith and jth body parts residing in P and Q.

Centroid Context-based Description of Postures

Posture Recognition Using the Skeleton and the Centroid Context T i is the ith normal behavior with the training threshold. q is the query posture. r i,j is the jth key posture of the ith normal behavior with length N.

Experiment Results