Abstract:
This work addresses the problem of anomaly detection in surveillance videos. To understand the challenges in this field, a comprehensive review of literature in the field was carried out. A suitable base system was selected from literature and analysed in depth. Then an approach utilizing the Histogram of Optical Flow (HOF) and Support Vector Data Description (SVDD) was proposed to overcome the shortcomings of the base system and improve its performance.
In the pre-processing stage, HOF was used to extract motion information (“events”) from video data. These events were then described using a compact feature vector, which encoded both spatial and temporal information. An SVDD, with a non-linear kernel for increased flexibility, then learnt a spherically shaped boundary around the dataset, which was then used to identify anomalous behaviour.
The performance of the proposed approach was evaluated on a publicly available benchmark dataset. The strengths of the approach are its flexibility in detecting a broad range of anomalies, its unsupervised learning method and its ability to learn complex non-linear motion patterns.
