Infrared Images-based Human Detection,Tracking And Discrimination Methods

Chinese Armed Police Force(CAPF)is responsible for the safety of the electric power plants(a million kilowatt or more)and the main bridges and tunnels of our country.If we concede the country as a human body,these facilities are like the heart and vessels to the country.There must be a huge number of property damage or even casualties if they were been destroyed.Violent terrorist activities against important targets mostly happen late in the night,the time that soldiers are tired and can't see clearly enough.Under the background of the country's mass disarmament,how to effectively reduce the duty forces occupied by these important targets and don't lose the protection of them is an urgent problem.Infrared radiation have longer wavelength than visible light,so it will be less affected by rain and fog.Video surveillance system based on infrared radiation is widely used in a variety of important places for security in terms its real time and all-weather monitoring.The approaches of human detection,tracking and classification based on infrared video surveillance were studied in this paper.The main contributions of this paper are as follows:1.The preprocessing method of infrared image is thoroughly studied.Several classical infrared image preprocessing methods are introduced in this paper,mainly including image denoising methods such as Median Filter and Gaussian Filter,as well as the morphological processing methods like dilatation,erosion,opening operation and closing operation.The denoising and morphological processing are both analyzed and tested based on the real infrared image data.2.Human detection methods in infrared images are researched.Several classical targets detection methods,including Frame Difference,Background Difference and Optical Flow,are introduced.On the basis of the theoretical research and the application of the targets detection algorithms,three-frame difference detecting method is realized and experimented based on a number of infrared images sequence.The results show that the human target can be effectively detected by three-frame difference method in infrared images.3.A novel human targets tracking algorithm,which is specifically aimed at infrared targets tracking is proposed.Three classical tracking algorithms,such as Particle Filter,Mean Shift and IVF(Intensity Variation Function)have been studied and analyzed.In order to solve the problems that exist in the classical algorithms,we propose a novel tracking method,which is called Mean Intensity Bordered Template Difference Algorithm.We use the target's template(the template's border has been replaced by the mean intensity of itself)of the previous frame subtracting the current frame,thus we can get the positive region and negative region.Then we calculate the vector between the centroid of the positive region and the negative region,after that we move the target's template along the direction of the vector with single pixel at one step,until the target's template approximately coincides with the human body in the current frame.The proposed method is more accurate and less time consuming compare with the Particle Filter tracking algorithm and Mean Shift tracking algorithm.Furthermore,we integrate the original tracking algorithm with adaptive tracking bonding box,pedestrian occlusion detection and the distinction after occlusion,which makes it more robust.4.An algorithm for classification of moving objects based on infrared image is improved.It's based on the original algorithm which was proposed by Cutler Ross.In this part,two approaches of the targets classification,one based on static features and the other based on dynamic features,were discussed.Eventually,we take the periodic features of the target as our approach.We select 100 continuous frames of the target(time span is about 3 seconds),and then calculate the targets' similarity of all the time delays so as to generate a similarity image.Then,the paper analyze the similarity image with four different measures such as one dimensional power spectrum,Fisher's Test,Time-Frequency Analysis and Autocorrelation Function.The results show that the Autocorrelation Function' feature is more prominent.So we use the Autocorrelation Function as the basis of target classification algorithm.The algorithm is tested on multiple infrared video data sets.The results show that the algorithm can effectively identify the moving human targets in infrared image sequences.