The large number of cameras present in public environments (i.e. airports, shopping centres) means that automatic solutions are required to monitor all incoming feeds. Recently, there has been an increasing interest in modelling activities in crowded scenes, and detecting abnormal events. Challenges of event detection in crowded scenes include occlusions, clustering and the background events coexisting with the event of interest.
In the learning and recognition part, we adopt two basic approaches for this problem:
- Unsupervised learning : assumes that the unusual events are those in low probabilities.
- Weakly supervised learning: assumes that we know the events if interest, and we have temporal annotation for the event in the training dataset but with the detailed location unknown.
In the feature extraction part, the following features are investigated:
- Optical flow
- LBP-TOP
- Particle Videos
- MPEG motion vector
- MPEG DCT coefficients.
PhD Student
