A novel fast bio-inspired feature for motion estimation

Motion features extracted from video streams are used in a wide variety of computer vision applications including action recognition. For this application many motion features are suggested in previous works and in order to improve the results a group of them are used with some classification methods. In this paper, based on a bio-inspired motion perception model in animals, a new motion feature is proposed. The model is simple and could be realized with a limited number of mathematical operations and the process of feature extraction is much faster than well-known techniques such as histogram of optical flow (HOF) or motion boundary histogram (MBH). Moreover, with proposed modifications the feature becomes invariant to pixels' brightness and mostly sensitive to the magnitude of the motion. Empirical results on KTH dataset show that this new feature outperforms many other typical features in action recognition and competes HOF with acceptable result of (94.49%), while being much faster due to its low complexity.