Research On Key Techniques Of Target Detection And Tracking In The Ground Scene

Target detection and tracking has always been a fundamental topic in computer vision and artificial intelligence.It provides the priori information for the subsequent tasks including pattern recognition and target behavior understanding.In recent years,the search and track technology for air objects has been developing quickly.However,the complexity of the ground scene,the diversity of ground objects,background cluster and noise seriously affects the precision and accuracy of target detection and tracking.How to suppress these interference factors in order to improve the target detection rate and reduce the false alarm rate(FAR)is an important task and urgent demand in real applications.In this paper,the interference factors,which may result in detection or tracking failure,are analyzed in the applications of target detection system.An in-depth research towards three topics is presented,including moving target detection with monocular mobile platform,target detection and tracking under complex background and camouflage target detection.A series of problems have been solved,which provides a strong reference to design the target detection and tracking system.In a mobile ground target detection system,the stationary objects with significant 3D structure in the scene will lead to strong parallax in the image due to the camera's ego-motion,which will result in a high FAR.The geometric transformation between the images is first derived based on the pin-hole camera projection model,and the target detection problem under a mobile platform can be divided into two catagories:the absence of parallax and the presence of parallax.According to the former,a method based on global motion compensation and motion history image is proposed.It utilizes the spatial correlation of the target contours in the image,and the motion history image is then weighted by successive images at different times.The target segmentation results are more precise and complete under this stategy.According to the latter,a series of constraints,including epipolar constraint,parallax consistency constraint and structure consistency constraint,are proposed based on the multiview geometry.An integrated framework of moving target detection and segmentation is established.The proposed method can overcome the interference of strong parallax and improve the detection accuracy.In the complex ground scene,the attributes such as illumination change,background motion and measurement noise can cause serious interference to motion target detection.In addition,the attributes including pose change,scale variation and occlusion tend to result in tracking failure.Given this,a robust principal component analysis(RPCA)framework that is superior to traditional target detection methods is introduced.Firstly,a total variation(TV)regularized RPCA method is proposed to handle the presence of background motion and detection noise.The foreground noise can be obviously eliminated and the FAR is reduced because TV method can suppress the noise while preserving the edge of target.Then,the instability of convex relaxation in the RPCA method is pointed out and a moving target detection algorithm based on the nonconvex proximal p-norm is proposed.This method can well constrain the sparsity of the foreground target and achieve a satisfactory result.Finally,a tracking algorithm based on contiguous occlusion constraint is proposed.The occlusion is described using a Markov Random Field(MRF)model and the complete target appearance is constructed.The whole tracking process is implemented in a particle filter framework,which makes the algorithm be robust to pose change,scale variation and illumination change.The occurrence of camouflage target interference is analyzed,when color and gray information of target and background can not be distinguished.We first introduce the representation and measurement theory of polarized lights and polarized optical systems.A target discrimination method based on the polarization characteristics is proposed.After that,the Mueller matrix of the target surface is obtained by the polarization measurement,and the optical constants of the material are derived to separate the target and the background.Then,the compressed Bidirectional Reflectance Distribution Function(BRDF)is extracted by projecting the multi-angle intensity information into a subspace.Based on the compressed BRDF features,the target and background pixels are clustered using the K-means classification method.Finally,a target detection method based on the Mueller matrix discrimination is proposed.The polarization difference image related to the polarization state of the optical system is computed,and of which the contrast between the target and the background is maximized using the Support Vector Machine(SVM).The polarization state of the detection system is adaptively determined in order to effectively discriminate the target from the clustered background.Experiments demonstrate that the proposed method can be well applied to the camouflage target detection.