Research on Foreground Object Recognition Tracking and Background Restoration in AIoT Era

摘要

With the rapid development of artificial intelligence technology and its combination with the Internet of things, recognition processing system has become a hot field of intelligent research. Meanwhile, with the popularity of SOC sensor, complex recognition algorithms are effectively implemented and utilized. In order to explore the characteristics and application scope of each foreground object tracking algorithm, this paper establishes foreground object tracking algorithms based on three kinds of Gaussian models and two kinds of Neural network models respectively, and solves the unique "delayed ghosting" phenomenon of Gaussian model with the total variational regularization (TV regularization). In the detection problem of camera jitter, a speed up robust feature (SURF) is added to optimize the model. On this basis, the kernel density estimation (KDE) is used for automatic threshold selection, and a secondary optimized foreground object recognition and tracking model is obtained. The results show that the recognition rate of the optimized mixture of Gaussian (MOG) model is more than 90% in static recognition, the recognition and tracking accuracy of the neural network model is higher than that of the Gaussian model in dynamic background, and the fuzzy self-organizing background subtraction (F-SOBS) can greatly overcome the camera jitter problem.