Mobile Object Detection with CNN and Multi-Timescale Background Models

The patent divides the background model into three distinct time-span models: short-term, intermediate-term, and long-term background models. Each model captures background characteristics at different temporal scales, allowing the system to segment the analysis of stationary objects across multiple time horizons. This segmentation enables the system to distinguish between temporary illumination changes and genuine stationary objects by comparing detections across the different time-span models.

The patent merges the results from three different background models with varying time spans to make the final detection decision. By combining the short-term, intermediate-term, and long-term background models, the system leverages the strengths of each model: the short-term model reduces false alarms from rapid illumination changes, while the long-term model captures stationary objects that persist beyond brief temporal fluctuations. The integration of these models resolves the contradiction between precision and reliability.

The patent introduces an intermediary mechanism by using the intermediate-term background model as a bridge between the short-term and long-term models. This intermediate model helps filter out the harmful effects of illumination fluctuation that affect the long-term model, while still maintaining the ability to detect stationary objects. The intermediate-term model acts as a mediator that smooths the transition between rapid changes and long-term trends, reducing the influence of external noise.