Top two rows: Frames from i-LIDS (AVSS2007) Parked Vehicle Challenge Dataset;
Bottom row: Identification of background, foreground, and midground at time 160.
Many crucial surveillance tasks require attention to new stationary objects that appear and persist over time in the midst of a rapidly changing, cluttered scene (e.g., a suitcase left unattended in a crowded airline terminal for several minutes, a person or vehicle loitering near a busy street). Salient objects in these scenarios cannot be classified solely as traditional foreground objects nor as permanent background features. Instead, they are part of a midground realm. Midground is defined by a temporal window following the object's appearance, but it also depends on adaptive background modeling to allow detection with scene variations, such as temporary occlusion and small illumination changes. The human visual system has evolved to detect rapidly moving foreground objects and is ill-suited to perceiving changes over the longer time scales characteristic of how midground objects appear. Automated video surveillance systems hold the potential to "tune in" to such changes within a specified temporal window.
This research introduces a midground detection technique which emphasizes computational and storage efficiency. The technique explicitly models a precisely defined temporal window during which foreground objects that have become stationary are classified as midground. This separates scene elements into three categories: long-lived, persistent background elements, short-lived (ephemeral), moving foreground elements, and newly stationary, persistent (non-ephemeral) midground objects. The approach uses a new adaptive, pixel-level modeling technique derived from existing backgrounding methods. Experimental results demonstrate that this technique can accurately and efficiently identify midground objects in real-world scenes, including PETS2006 and i-LIDS (AVSS2007) challenge datasets.
|
 |
 |
 |
Publications: