Attention computation model for coal-mine surveillance video based on non-uniform sampling in spatial domain and time domain

There are few methods to deal with the coal-mine surveillance video, the only few existing methods all make use of the traditional image engineering ideas. Visual attention has the prominent functions that can reduce computation and accelerate the computing speed. This paper firstly analyzes the limitations of existing down-top attention model in the application of coal-mine surveillance video, then proposes a new visual attention computation model based on sequential scale space and multi-features. Different from the existing model, non-uniform sampling in spatial domain of our algorithm is expressed as discrete structure of sequential scale space, and we choose the modified Bessel function as the smooth kernel, in time domain, we establish a threshold for frame sampling. About feature extraction, we choose the motion conspicuity, wavelet package decomposition and gray intensity as measures of saliency, DOG (Difference of Gaussian) operator as the generalized method. Finally, a global saliency map for the interesting objects is formed. The experiment results show the flexibility and effectiveness of this model.