Anomaly detection in videosurveillance feeds
Traditional passive surveillance is proving ineffective as the number of available cameras for an operator often exceeds the operators ability to monitor them. Furthermore, monitoring surveillance cameras requires a focus that operators can only uphold for a short amount of time. Algorithms for automatic detection of anomalies in video surveillance feeds are thereby constructed and presented in this thesis by using hidden Markov models (HMM) and a Gaussian mixture probability hypothesis density (GM-PHD) filter. Four different models are created and evaluated using the PETS2009 dataset and a simulated dataset from FOI. The three first models are created to model normal behaviour of crowds in order to detect anomalies. The first uses only one HMM to model all observed behaviours. The second model uses two different HMMs, created by manually splitting the observations in the training set into two parts corresponding to different behaviours. This model does not perform as well as the first model. The third model is attained by clustering the observations in the training dataset, using dynamic time warping (DTW) and z-scores, and creating a separate HMM for each cluster. This model is regarded as the most efficient anomaly detector. The last model uses information from all crowds in the surveilled scene but does not perform well enough to be used to detect anomalies.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)