This paper proposed a new background model for motion detection in video surveillance based on neural network (NN).A neural network background model was build for every pixel (or a small local region).It is a four-layer feedforward neural network.Input layer accept HSV pixel value

feature layer extract features form HSV

pattern layer work as a background probability calculator.Output layer classifies the pixel into background or foreground

and finds the activated node.Weights and structure of network updated dynamically along with motion detection and no training video needed.Adaptability of background model includes adaptive learning rate calculated form motion difference between adjacent frames

and number of pattern node changes according to weight variation.Experimental results on benchmark videos show that

without any manual setting of learning rate

the proposed algorithm can detection motion more precisely than other familiar background models

and it can also adapts to sudden background or lighting changes more quickly.