Optical flow-based enhancement of spatio-temporal detection in videos

Abstract

Accurate optical flow techniques are widely used in spatio-temporal object detection in videos. However, the computational complexity of the currently used techniques limits the effectiveness of spatio-temporal detection in applications such as action detection and event recognition. Therefore, in this paper we aim at employing rapid yet accurate optical flow techniques to promote the effectiveness of the detection system. The proposed design uses novel optical flow estimation techniques that are based on learned flow basis, known as PCA-Flow and PCA-Layers. PCA-Flow estimates dense flow from a linear flow model based on principle components of natural flow. PCA-Layers is an extension of PCA-Flow. PCA-Layers technique uses Markov random field (MRF) to combine several motion layers into dense optical flow. The motion in each layer is estimated by PCA-Flow. Our experimental results show that our approach maintains the overall performance of the baseline framework while 64% reduction in the computation time is achieved.