A saliency based motion detection model of visual system considering visual adaptation properties

Abstract

The purpose of this study is to construct a mathematical model which predicts saliency regions in high-speed egocentric-motion movies, filmed by an embedded camera in a driving vehicle, by reproducing the characteristics of the area MT and MST neurons' receptive fields with consideration of visual adaptation properties. The area MT neurons integrate from the area V1 activation and respond well to regions where higher motion contrasts exist. While the area MST neurons detect global motions such as expansion, contraction, rotation, and so on. We modeled the area MT neurons' receptive fields as a center-surround spatial summation of counter sided motion vectors of visual scenery. The area MST neurons in our model integrate the responses of the MT neurons by convolving with spacial weight functions of which central portions are biased to preferred direction. Visual adaptations were taken as the primary delay filters for each visual feature channel to deplete the saliency of stationary objects and regions during particular frames. The simulation results for the movies which were taken in a running vehicle indicate that the proposed model detects more salient objects around the vanishing point than the conventional saliency based model. To evaluate the performance of proposed model, we defined the moving-NSS (normalized scan-path salience) scores as the averaged NSS scores in each moving time window. The moving-NSS scores for motion images of our model were higher than those of the conventional model.