Research On Infrared Human Detection And Tracking From Complex Background

The development of society and economy drives the advancement of technology,and so infrared technology does.Compared with video surveillance system(VSS)based on visible light,system based on infrared could still work in bad weather like foggy,smokey and rainy as usual.With respect to the improvement of the automation of VSS,computer vision provides us with an excellent solving scheme and guides the developmental direction.However,due to some factors like low signal noise ratio(SNR),lack of information and non rigidity of human,infrared human detection and tracking have always been a hot and difficult point in the field of computer vision.This paper is mainly about infrared human detection and tracking in complex background,and the main research works are as follows:1.We present an algorithm of motion object segmentation based on modified Gaussian Mixture Model(GMM).And the candidate objects are further segmented based on graph cut(GC)and active contour model(ACM).Although GC and ACM are usually utilized in segmentation of medical images,we extend its application.Combined with GMM,a more accurate object is obtained by local segmentation.The results of different frames indicate that the accuracy of segmentation of the algorithm is higher than others.2.For the purpose of detecting pedestrians from outdoor surveillant video,a feature called entropy-edge weighted local gradient orientation descriptor is proposed.Firstly,the magnitude and orientation of gradient of every pixel are computed and the orientation is divided into 9 bins.Then the image is separated from many blocks in accordance with histogram of oriented gradient(HOG).After that,we sum the gradient in the same range in every block and the summation is mapped to the corresponding pixel.That is to say,every block represents a pixel and as for every bin,we could get an image called “orientation image”,so 9 images can be generated in all.The value of every pixel is the sum of the gradient with the same orientation range in the corresponding block.In addition,in order to enhance the effect of edge,we fuse the feature with edge histogram within 5 directions to get a better robustness.At last,the feature vector is normalized by different normalization factors because of the different concept of gradient and edge.Cross validation and experiments on other testing videos shows that the accuracy and robustness of the proposed method are better than other feature like HOG.3.Human tracking is also studied in this paper.Based on the special characteristics of infrared imaging and human object,a joint model based on spatial information and kernel histogram is proposed.What's more,the bandwidth of kernel function is updated on the foundation of improved GMM.Firstly,there is little information which could be extracted from infrared image.In addition,the tracking process is far from robust if it is only based on grayscale.For the purpose of robustly tracking,an improved target model with spatial and grayscale information is consequently established.Considering that the bandwidth of kernel function is a constant,a rough region where moving target may appear is located by improved GMM.The optimal location of target is searched in this region iteratively.Experiments indicate that the model proposed is robust to some extent and the updating mechanism of the bandwidth could accommodate the change of the size of human.