Temporal weighted learning model for background estimation with an automatic re-initialization stage and adaptive parameters update

Background initialization and background update are two important stages considered in the design of most background modeling algorithms. Commonly, these algorithms implement strategies in which their parameters have a very high adaptability in the background initialization stage in order to learn all the variations of the background. Contrary, in the background update phase, these parameters adapt slowly, in most cases with an exponential decay. This paper presents the BE-AAPSA method which automatically determines if the background initialization and the background update need to be re-initialized. Re-initialization is triggered if the video scene presents high variations, allowing the background to be defined more accurately. BE-AAPSA is based on a previously developed system, where two adaptive background models based on weight arrays with temporal learning mechanism identify dynamic objects within a video scene. The system implements four independent modules to treat the different factors that affect a correct definition of the dynamic object. In BE-AAPSA, the objective is to create a robust background estimation model where the learning rates for each pixel are calculated according to the results of the two adaptive weight arrays and the module where the video is classified. This approach allows handling different strategies to update learning rates at a pixel resolution. BE-AAPSA is validated with the SBI and SBMnet video databases and with a video created by concatenating scenes of the video categories presented in the CDnet 2014 database. According to the findings, BE-AAPSA produced highly accurate results with SBI and SBMnet and surpassed state-of-the-art methods with the CDnet video. These results demonstrate the importance of using an automatic re-initialization scheme in the background initialization and background update stages when the video scene presents a major change or involves jittering. Furthermore, it shows the benefits of handling in separate modules the analysis of the background estimation results.