Action Recognition Research With Analog Model Of Neurons In Primate Visual Cortex

Recognizing human actions from videos is important to video surveillance, videoretrieval and human-computer interaction. However, it is quite difficult to recognizeaction rapidly and accurately, due to the highly complicated human posture andanthropometric differences between individuals. With the rapid development ofcognitive sciences and neuroscience, image processing by modeling visual perceptionmechanism attracts the attention of researchers, especially in the human actionrecognition. During the past four decades, hundreds of recognition models have beenproposed in the1iterature.But, those models usually suffer from less biologicalplausibility. Therefore, in order to improve the accuracy of human action recognition,accelerate the recognition speed and explore the characteristics of visual perception, wepropose a method for human action recognition by modeling the visual perceptionsystem.Physiology reveals that the visual cortex is organized in two different pathways:ventral stream and dorsal stream. The dorsal stream involves in the analysis of motioninformation with areas V1, MT. Although there are several theories which speculate onhow and where pattern motion is computed from V1outputs. Up to now, a large part ofthe models center round MT. However, those models are very complex and none ofthem has been tested in a real application for action recognition. In this paper, the authorinvestigates the pattern motion is computed from V1, mainly done the follow works:Firstly, base on the theory of marr, this paper gave a rapid method of human actionrecognition. Meanwhile, it simulate the classical receptive field （CRF） of simple andcomplex cells in the primary visual cortex with3DGabor spatial-temporal filter and itscombination to process the video sequence, in order to obtain the sport energy that issensitive for the sport speed and direction.Secondly, according to surround inhibition benefit in motion detection and reducethe texture. This paper propose surround inhibition model with direction and nodirection in V1, it enhanced the sport energy and reduced the influence of noise.Finally, due to visual perception system with spiking as the information processingand transmission carrier, conductance-driven integrate and fire neuron model was usedto simulate the primary visual cortex neuron, by which motion information wasconverted into spike train. And the mean firing rate of spike train formed a featurevector that captures the characteristic of human actions in this video sequence.The method is tested on the Weizmann and KTH action dataset. The obtained impressive results showed that the pattern motion is computed from V1.