Multi-feature Fusion Target Tracking Based On Particle Filter In Complex Scene

Target tracking technology is to detect moving targets in video image in order to capture the target motion information, such as real time location, movement trend, tracking,etc. It is widely used in intelligent surveillance, visual navigation, intelligent transportation,human-computer interaction, defense reconnaissance and so on. Particle filter algorithm is a statistical filtering method based on monte carlo and recursive bayesian estimation, has shown its unique advantage in dealing with parameter estimation and state of filtering in the system of non-linear and non-gaussian. In recent years, target tracking technology based on particle filter has become one of research hotspots in the field of machine vision.But it still has some problems in automatically perceiving and capturing the moving target,target easily lost, tracking accuracy and stability of poor in practical applications due to the influence of the factors in complex scene, such as inaccuracy of target template acquired in the initial tracking stage, target color similar to the background color, attitude change,occlusion, camera shake, multi-target tracking and etc. In order to improve the accuracy and robustness of the particle filter target tracking in the complex scene, this thesis based on the actual application demands of research projects is devoted to using the theory of particle filter target tracking framework to explore the structural target characteristic expression model, multi-feature fusion rules and adaptive target detection method. The main contents and contribution of this dissertation are as follows:(1)Research on the defects of traditional particle filter target tracking algorithm.In practical application, the traditional particle filter algorithm has two main defects:First, the traditional particle filter algorithm can’t real-time, automatic perception and capture the moving targets in the scene by manually locking the first frame video target,especially for multiple targets tracking task, It can’t effectively perceive targets that will lead to can’t achieve targets tracking task when different targets in the scene appeared frequently, Second, the traditional particle filter algorithm using a single RGB color histogram as the reference target probability model and the candidate target probability model, then the bhattacharyya distance is used to calculate the similarity degree between the reference target probability model and the candidate target probability model, and according to similarity degree to updating particle weights, and the re-sampling particle weight and weighted criteria is used to estimate the target state location information.In this process, because the three components of the RGB color space are not independent,the target is easily lost when on the condition of the light change or target color similar to the background color, At the same time, a single RGB color features do not have express target geometry information, the target tracking also can be easily lost when on the condition of the target is partly occluded or camera motion imaging complex scenes.(2)Analyses the complexity of the video scene and Research on the influence of particle filter target tracking performance in different scenarios with different target features model.The validity and accuracy of particle filter target tracking algorithm depend on how to establish target feature model according to the characteristics of the target scene. First,this thesis analyzes the complexity of video scene, which include the light change, target color similar to the background color, the target geometrical deformation, occlusion,camera shaking, multi-target tracking and so on; Then study the effect of particle filter target tracking performance in different scenarios with different target features model, and adopt the RGB color model, HSV color model, the LBP(Local Binary Pattern) texture model, HLBP(Haar LBP) texture model and SUSAN corner point model on six different scene video of experiments were carried out respectively. The experimental results show that particle filter target tracking algorithm based on color has good tracking robustness because color feature model is not sensitive to target attitude change or geometric deformation, and particle filter algorithm based on texture feature has good tracking robustness when on the condition of target color similar to the background color, partial occlusion or camera shake because texture features model has a good illumination invariant and can reflect the target area like yuan structure information, and particle filter algorithm based on corner feature has good tracking robustness when on the condition of target color similar to the background color, the target geometrical deformation, partial occlusion and so on because corner is significant feature of the target contour which is not affected by target deformation, and with rotation invariance, translation invariance and scaling invariance.(3)Sobel median binary pattern(SMBP) texture description operator with edge information and light resistance is proposed in this paper. And then,the HSV color kernel function histogram probability model and SMBP texture histogram probability model are fused. On the basis, multi-feature fusion single target tracking algorithm based on particle filter in complex scene is proposed.The sobel median binary pattern(SMBP) texture description operator is proposed,the operator for sobel edge detection and median filter in the LBP window target area, It not only inherited the rotation invariance and the gray scale invariance of the LBP texture operator, and can effectively capture the texture feature of target area, but also have to capture the target edge information and the ability to resist noise; Accuracy of the particle filter target tracking based on SMBP texture feature algorithm has improved significantly in the condition of target color similar to the background color, partial occlusion,camera shake and so on. On this basis, according to the target tracking application requirement of complex scenes, such as the illumination change, attitude change, occlusion and so on.study the fusion rule between the HSV color kernel function histogram probability model and SMBP texture histogram probability model to proposes feature fusion target tracking Based on particle filter algorithm in complex scene. The experimental results show that this algorithm has good tracking accuracy and robustness.(4)Multi-feature fusion multi-target tracking based on particle filter algorithm in complex scene is proposed.According to the demand of multi-target tracking test, this paper puts forward a kind of multi-target automatic detection algorithm that automatic perception and capture moving target. This algorithm using the background difference method and symmetric difference method to detect target area in the second frame of video image respectively,and then using the maximum entropy segmentation method to determine the threshold of the detection after image and binarization processing, and get the result mask calculation to get more accurate binarization images of every target region; Then through the connected domain calibration algorithm determine center position of each target area, thus, realize the automatic perception and capture of multi-objective. On this basis, multi-feature fusion multi-target tracking based on particle filter algorithm in Complex Scene is proposed. The experimental results show that the multiple target tracking algorithm can effectively automatic test the position and number of moving targets in complex scene, and has excellent tracking robustness.