Change Detection Algorithms

Within the field of Computer Vision, in the last decade a wide range of research areas concerned with real-time applications have received a growth in attention, due to a considerable performance boost of off-the-shelf computing platforms. This is particularly true for those fields paving the way for emerging applications in unconstrained environments wherein, unlike established industrial application, a complex and changing world must be accurately and reliably modeled. One of these research fields is undoubtedly change detection.

Change detection deals with the automatic detection of “significant” changes occurring in a scene by the elaboration of single or multiple video sequences captured from single or multiple view-points by fixed or moving imaging devices. Change detection is the first crucial processing step in many Computer Vision applications, such as video-surveillance, traffic monitoring and remote sensing. In fact, upon a reliable preliminary change detection step higher level capabilities can be built, such as those concerned with objects tracking, classification and behavior analysis.

Changes are regarded as being “significant” if they correspond to variations of the imaged scene geometry. Actually, not all measurable image changes can be ascribed to geometrical scene changes, for possible disturbance factors may induce “false” changes. The main disturbance factors acting in real-world applications are:

  • camera noise
  • dynamic adjustments of the camera parameters (e.g. auto-exposure, auto-gain)
  • scene illumination changes:
    • global illumination changes (e.g. a cloud passing by the sun, light switches)
    • local illumination changes (e.g. shadows cast by moving objects, light spots)

We have devised different change detection algorithms, each of them adressing specific issues and to a large extent complementary so as to allow for integration into a global approach. Mainly, the approach relies on a single-view change detector robust to camera noise, dynamic adjustments of camera parameters and global scene illumination changes, together with a multi-view information fusion paradigm aimed at dealing with local illumination changes.

Single-View Change Detection

Multi-View Information Fusion