Anomaly Detection for Video Surveillance Applications Carmen E. Au, Sandra Skaff, and James J. Clark McGill University, Montreal, Canada Abstract We investigate the problem of anomaly detection for video surveillance applications. In our approach, we use a compression-based similarity measure to determine similarity between images in a video sequence. Images that are sufficiently dissimilar are deemed anomalous and stored to be compared against subsequent images in the sequence. The goal of our research is two-fold; in addition to detecting anomalous images, the issue of heavy computational and storage resource demands is addressed. 1. Introduction In this paper, we consider the problem of detecting anomalous or novel images in a sequence of images. The need for heightened security is unfortunately becoming more prevalent in today’s world. While aptly-placed video cameras are intended to capture the images of possible offenses or offenders, the majority of the time, it is but a security guard who spends his or her time staring into a series of uneventful monotonous images. Since new events rarely occur, it is extremely difficult for a security guard to remain vigilant at all times. Thus, our work removes the onus of detecting anomalous situations from the security guard; and rather, places it on the video surveillance system. We developed a compression-based similarity measure technique, which compares the sizes of compressed images to detect for anomaly. While there exist other approaches to automated video surveillance [3, 4, 7], our compression-based technique inherently provides storage and computational resource reduction. One way of detecting novel images is to compare each new image to all the previously-seen images. However, this method is not ideal; storing every image of the video sequence requires significant memory and far too many unnecessary comparisons. Thus, by storing only the novel images, it is only the images with useful information that are kept and used to be compared against future incoming images. 2. The Similarity Measure A scene is considered anomalous when the maximum similarity between the scene and all previously viewed scenes is below a given threshold. We propose to use a similarity measure that quantifies the mutual information between data sets. By definition, compression programs aim to remove redundancies within a file. Li et al [5] proposed a similarity metric that involved compressing each data set independently, measuring the sizes of the compressed results and then concatenating the two data sets and compressing the combination data set. The idea is that the size of the compressed version of the concatenation of two similar files would be smaller than that of two dissimilar files. In terms of the normalized compression distance (NCD) developed by Li et al, our similarity function is given by: [ ] ( ) [ ] ( ) [ ] ( ) [ ] ( ) [ ] ( ) 1 2 1 2 1 2 1 2 ( , ) size C I size C I size C I I pI I size C I size C I + − ⊕ = + where C[] indicates the compression operation, and ⊕ indicates the concatenation of the data sets. Though this compression-based technique discards much information, because we are merely detecting anomalous images instead of attempting to detect and recognize the nature of the dissimilarity, we are not concerned with the discarded information. In fact, since our system need only be able to say that it has seen a similar image before, and does not need to say what is similar, we could even implement a particularly lossy compression technique which throws away much information that would otherwise be needed to recognize what is dissimilar. However, not all types of compression algorithms are suitable for implementing such a similarity measure. We require an algorithm that can detect and remove the redundancy in a concatenation of two images. Block-sorting lossless algorithms, such as those derived from the Burrows- Wheeler transform [1], for example the bzip2 algorithm, are well-suited to our needs.. By default in the bzip2 program, files are compressed in 900 KB (1) The 18th International Conference on Pattern Recognition (ICPR'06) 0-7695-2521-0/06 $20.00 © 2006