| With the development of society and economy, the requirement for security alertsystem are enhanced in all fields, and as an important means of security alert, videosurveillance system based on visible optical has been widely used. But due to theinherent flaws of such systems, it cannot meet the market demand. First, the commonused monitoring system is not intelligent enough, and it requires a lot of auxiliarymonitor by the aid of security personnel, it lacks of real-time. Second, in the case ofpoor visual environment, monitoring system based on visible optical cannot worknormally. Therefore, the infrared-based intelligent alert system, in recent years, isgaining more and more extensive research, but the related algorithms of human bodytarget detection, tracking and identification in infrared images are still not matureenough. In this paper, human detection and identification algorithms in infrared alertsystem are studied. The main research results in this paper are as follows:(1) To enhance the images and suppress noise in the images of the infrared alertsystem, this paper used a fast image enhancement algorithm; and for the problem thatthere is a deviation of translation in the horizontal direction between the two adjacentframes, this paper used a local projection matching algorithm. The fast imageenhancement algorithm proposed in this paper, fully absorbed the merit of selectedneighborhood image averaging algorithm and the characteristics of fixed relative ofthe backgrounds in the video sequence images, set the average gray of the imagesunder different temperatures as a standard factor, and regard the ratio of mean of thesingle frame image to the standard factor as weights of each pixels to smooth theimages, so that enhance the contrast of the images, and to reduce the interference ofnoise, instead the "bright spots", which three times higher than the standard imagefactor, of the mean of the image. The experiments show that this simple imageenhancement method, based on the imaging characteristics of the system itself,reduced the image noise and enhanced the contrast of the infrared image. The wholepixels of the images will be horizontally shifted because of dithering of the detectorwhen it is running; it needs to be registration before target detection. In view of theimage is shift in only one direction, this paper used a matching method of vehiclerejection, besides we used local rejection to reduce the effect of the moving objects toimage matching. The experiments show that the algorithm proposed in this paper isfaster and better than the classical template matching algorithm.(2) In this paper, using a background subtraction method to detect the moving targets in infrared video sequences. It needs to initialize the background beforeestablishing the background model, this paper used the single frame image which hasbeen eliminated the targets as the initial background. For the ROI extraction algorithmbased on single frame image, a bidirectional projection ROI segmentation algorithmbased on an adaptive threshold algorithm using scoring method and another ROIsegmentation algorithm based on the FAST corner characteristics and the CS-LBPtexture features are proposed in this paper. And for the algorithm based sequenceimages, this paper presented a selective grading adaptive background updatealgorithm. In this paper, experiments were conducted to verify the proposed algorithmThe experimental results show that the ROI extraction algorithm based on thesingle-frame image can extract human target areas and areas similar to the humantarget feature quickly and accurately; the algorithm based on sequence imagesproposed in this paper, overcome the disadvantage of the classical adaptivebackground modeling algorithm, that it will fetch in “ghost†when the backgroundmoves or the moving objects in the image are too slow, effectively. And theexperiments show that the selective grading adaptive background update algorithmcan accurately build the background model and extract moving targets in the images.(3) For the issue of ROIs classification and identification, this paper gave fullconsideration to the shortcomings of insufficient sample data, presented amulti-feature parallel extract method for ROIs classification and recognition. Thispaper presented a classifier based on improved Haar-like feature extraction algorithmsand a multi-resolution center-symmetric local binary pattern (MR-LBP) featureextraction algorithm. And selected the Adaboost classifier and SVM classifier, andtook full advantage of the feature extraction algorithm presented, and made up for theshortcomings of insufficient sample space, it can accurately divide the ROIs into threecategories, human target, cars and other non-human target. The experiments show thatthe target detection probability of the infrared alert system designed in this paper ismore than95%and the probability of false alarm is less than5%. |
