Flame Target Detection Based On Binocular Stereo Vision

Fire detection based on visual technology has become a hot research topic in recent years,and flame detection is the key technology in visual fire detection.Although scholars at home and abroad have put forward some algorithms about flame detection and achieved success,there are still shortcomings of poor robustness and low recognition rate in complex environment,and it is difficult to achieve satisfactory accuracy in flame location,which seriously restricts the further development of the flame detection of visual image.Therefore,the recognition and location of flame based on visual image in complex outdoor environment is a significant work.To solve the above problems,this paper studies the matching method based on the binocular vision model,and applies the human visual selection attention mechanism to establish the identification model of the static flame.An incremental support vector machine algorithm is proposed to identify the dynamic flame.The specific work of the paper is as follows:(1)In order to optimize the accuracy and speed of binocular stereo matching algorithm,adaptive window and weight stereo matching algorithm are studied in the paper.Since the information in the edge area of the image is rich,adaptive window algorithm based on edge is studied.Adaptive weight algorithm based on row and row parallel accumulation is used to reduce calculation quantity for non-edge regions.Finally,the position of the flame is obtained according to the disparity map and the camera intrinsic and extrinsic parameters.(2)In order to improve the discriminant accuracy of the static flame image,a static flame recognition model based on probability and visual salience frequency tuning is established according to the human visual attention mechanism in this paper.By analyzing the flame color value,luminance value and visual salience,we find that the model of visual salience can be used to recognize the flame effectively when the main color in the whole image is green or blue,otherwise,the flame is recognized by the probability model based on the color value and the luminance value.On this basis,the segmentation of multiple fire targets in a fire scene is studied.(3)In order to distinguish between the suspected flame target and the flame target in video,this paper studies the method based on five frame difference and Surendra background difference to extract moving targets from video sequence and get the intersection of moving target images and frame images,then analyzed eight characteristics of the dynamic flame and completed the definition and quantification of these characteristics.Then these characteristics are used as input feature vectors to train an incremental support vector machine classifier.Finally,the paper studies an adaptive genetic algorithm for optimizing parameters of incremental support vector machines,which effectively solves the problem that the classification rate is low due to the additional new samples.The simulation results show that the recognition rate of static flame based on probability and visual salience is 93.8%,and the error recognition rate is 8.6%.;The average recognition rate of dynamic flame based on incremental support vector machine is 95.5%and the error average recognition rate is only 3.2%;The positioning accuracy based on binocular vision model is less than 10cm.The simulation experiment results show that the static flame recognition model based on the probability and visual salience has high recognition rate and strong robustness.The dynamic flame model based on the incremental support vector machine can distinguish between the flame targets and suspected flame targets effectively.The matching algorithm based on adaptive window and weight can locate the flame effectively.It provides a basis for judging the trend of flame propagation and controlling fire.