One interesting application of midground detection is monitoring vehicle traffic and signaling an alarm when a vehicle parks illegally. In this application, vehicles must be detected that drive into a scene, stop, and remain stationary in no-parking zones, such as double-parking on a heavily trafficked street or parking partially on the street but with one or more wheels on the curb.

Scenes from the i-LIDS (AVSS2007) Parked Vehicle Challenge Dataset at each of the
three difficulty levels - Easy (a), Medium (b), and Hard (c).

Our approach to detecting a common class of illegal parking behavior is to combine midground object detection with activity region identification. Illegally parked vehicles are often midground objects (i.e., moving foreground objects that stop and remain stationary for a given period of time before either moving again or migrating into the scene's background). In our application, we identify midground objects that are stationed in no-parking zones that coincide or overlap with high-traffic areas and that satisfy appearance constraints characteristic of vehicles in a given depth of field. To flexibly and automatically identify this class of no-parking zones across widely differing scenes and in the presence of camera jitter, we identify areas of the scene, called activity regions, containing a relatively high frequency of foreground objects. We are experimenting with a set of real-world video sequences from the i-LIDS (AVSS2007) Parked Vehicle challenge dataset.