An Efficient Spatiotemporal Approach for Moving Object Detection in Dynamic Scenes

The dynamic texture (DT) which treats the transient video process a sample from the spatiotemporal model, has shown the surprising performance for moving objects detection in the scenes with the background motions (e.g., swaying branches, falling snow, waving water). However, DT parameters estimation is based on batch-PCA, which is a computationally inefficient method for high-dimensional vectors. Besides, in the realm of DT, the dimension of state space is given or set experimentally. In this work, the authors present a new framework to address these issues. First, they introduce a two-step method, which combines batch-PCA and the increment PCA (IPCA) to estimate the DT parameters in a micro video element (MVE) group. The parameters of the first DT are learned with the batch-PCA as the basis parameters. Parameters of the remaining DTs are estimated by IPCA with the basis parameters and the arriving observation vectors. Second, inspired by the concept of “Observability” from the control theory, the authors extend an adaptive method for salient motion detection according to the increment of singular entropy (ISE). The proposed scheme is tested in various scenes. Its computational efficiency outperforms the state-of-the-art methods and the Equal Error Rate (EER) is lower than other methods.