Overview

We propose a low level method for learning pattern of activities and detecting abnormal events. We proceed directly at the pixel(s) level, based on motion labels obtained with background subtraction.

The main advantages of our method are:

• the abnormal activity is NOT known a priori,

• we do NOT rely on object extraction and object tracking.

For one (or many) key pixel(s), we learn co-occurrence statistics of motion labels in the spatio-temporal neighbourhood of key-pixel(s). This learned normal co-occurrence distribution is then used for abnormal events detection.

For more details, please refer to the CVPR'09 paper.

Some experimental results

The first example (see Fig. 1) is one which shows normal traffic and cars making illegal U-turns. As shown in Fig.1(c), the trace left by the U-turn significantly differs from the usual traffic flow model. We can observed that the co-occurrence matrix contains information about the regular traffic flow but also activities generated by pedestrians crossing the street.

(a) Input video (b) the co-occurrence matrix of regular traffic flow and pedestrians crossing the street (c) the trace left by a car making an illegal u-turn.

The second example shows a person dropping a baggage and abandoning it. In this video, pedestrians usually walk from left to right and from right to left, hence the X shape of the co-occurrence matrix (see Fig. 2 (b)). When the person drops the bag, the abandoned package leaves a straight elongated line (see Fig. 2 (c)) which differs from the co-occurrence matrix and thus causes this situation to be suspicious.

(a) Input video (b) the co-occurrence matrix modeling pedestrians walking from left to right and from right to left (c) the trace left by a person dropping a bag.

The dataset can be downloaded here !

Y. Benezeth, P.-M. Jodoin, V. Saligrama, C. Rosenberger, Abnormal events detection based on spatio-temporal co-occurences, IEEE international Conference on Computer Vision and Pattern Recognition (CVPR), 2009. (pdf) (poster)