The goal of the project is to develop robust techniques for the recognition of unexpected behavior of crowds and individuals in complex situations. The goal encloses the following aims: a. Combine statistical features with tracking of individuals. Statistical features are more robust for a global decision of unexpected behavior with the drawback that no information of individuals is used. Global and local approaches can benefit from a combination of both. b. Improve computational performance of people detectors such as the detection using the histogram of oriented gradients (HOG) by learning spatial relationships in the scene in order to execute in real-time on smart cameras. c. Improve trajectory assignments for tracking using discriminative distance functions. d. Extend the two dimensional HOG model or another model to a three dimensional representation using a multi-camera setup. The solution of the behavior recognition problem is important in many visual surveillance scenarios. Examples are: o Bank robbery: Individuals who behave unexpected can be tracked and recorded. A possibility would be that the system locks all doors and calls the police when unexpected behavior is detected. o Crowd analysis: General crowd analysis includes events such as evacuation, mass panic and gathering. Behavior recognition can be applied to detect unwanted events.