Image Processing System for Stationary Object Detection Reliability

The patent segments the detection process into two distinct stages: first comparing short-term background models to detect moving objects, then comparing long-term background models to detect left-behind objects. This segmentation allows the system to handle dense moving object flows in the first stage and stationary object detection in the second stage, preventing false detections by processing different object types separately rather than simultaneously

The patent introduces a moving object detection result as an intermediary element that mediates between the short-term and long-term background model comparisons. By using the moving object detection results to mask or exclude regions with dense moving object flows before performing long-term background model comparison, the system prevents these flowing regions from causing false detections in the left-behind object detection process

The patent implements dynamic background model updates with different time constants for different detection purposes. The short-term background model uses a faster update rate to track recent scene changes and moving object flows, while the long-term background model uses a slower update rate to maintain accurate stationary background representation. This dynamic updating strategy allows the system to adapt to changing scenes without causing moving objects to permanently blend into the background

The patent performs preliminary detection of moving objects using the short-term background model before the long-term background model comparison is performed. This preliminary action identifies regions where moving objects are present, allowing the system to exclude these regions from the long-term background model update process, thereby preventing moving object information from being lost or blended into the long-term background model