Video smoke detection has many advantages over traditional methods

such as fast response

non contact.But most of current methods for video smoke detection have high rates of false alarms.Through analyzing the characteristics of smoke motion

a novel video smoke detection is presented.In order to accelerate detection speed

video images are divided into blocks.Each block motion orientation is estimated by block matching methods.And a time sequence of motion orientation for each block is generated over a sliding time window.Then accumulation and main motion orientation are computed according to the sequence.The accumulation represents the degree of motion duration and the main motion orientation describes the maximum possible orientation of each block over the time window.A 3D feature is extracted from the accumulation and main motion orientation

and a Bayesian classifier is used for smoke detection.Experiments show that the algorithm is robust and significant for improving the accuracy of smoke detection.