Research On Infrared Pedestrian Target Detection And Tracking Technology In Complex Scenes

As the cost of infrared sensors decreases,intelligent monitoring system based on infrared imaging devices is becoming more and more popular.The main functions of intelligent monitoring system are detection,tracking,identification,behavior understanding and so on.Among them,target detection and target tracking are the important technology of the system front-end function.However,due to the complexity of the ground environment and the variability of pedestrian targets,the detection and tracking of pedestrian targets in infrared images are still challenged.This paper mainly studies from the following aspects:According to the characteristics of infrared images,the target characteristic,background characteristic and noise characteristic of the infrared image are studied respectively,which provides the theoretical basis for the following research on image preprocessing,pedestrian detection and tracking.For the infrared image prepossessing,the non-uniformity correction algorithm is studied in image noise suppression.This paper focuses on the least mean square error method and constant statistic method in the scene correction method.In view of the ghosting phenomenon,an adaptive mean filter correction algorithm based on constant statistics is proposed.The mean and offset of each detection unit are expressed as the mean filter of parameter correction.The optimal correction parameters are selected through variance estimation method.The results show that the ghost phenomenon can be effectively suppressed.Aiming at the problem of pedestrian detection in infrared images,firstly,based on the analysis of common methods,this paper focuses on the extraction of candidate regions based on Gaussian mixture model.For the low background update,which is caused by the constant learning rate in a certain case of the number of Gaussian functions.This paper proposes a method based on motion history image and mixed Gaussian model.This method can dynamically adjust the update rate to make the background model more adaptable.The results show that the proposed algorithm can achieve good detection results.Secondly,the idea of the global to the local is used,and the target representation is implemented.Finally,the pedestrian classification is completed by using the SVM training classifier.Through the experiments on different scenes,the results show that the pedestrian targets can be detected effectively.For the problem of pedestrian tracking in infrared images,the particle filter theory is studied in this paper.Based on the previous research,the target observation model is constructed by using local features(HOG and LDP).On account of the model,a weighted fusion method based on the Bhattacharyya coefficient and a target model updating algorithm are proposed.Finally,a particle filter infrared pedestrian tracking framework based on adaptive multi-feature fusion is established to achieve target tracking.The results show that the algorithm can effectively track the target,and the algorithm has strong robustness in a certain occlusion and adhesion.