Research On Pedestrian Detection And Tracking Method Based On Infrared Video

With the continuous development of infrared thermal technology,the resolution is constantly increasing and the demand for nighttime monitoring is increasing.No matter in daily life or in the military,the application of infrared technology is increasing.At night,infrared thermal video is more efficient than visible video,so the detection and tracking of targets in infrared thermal video has also been studied by more and more people.however,due to the characteristics of infrared thermal video,it has brought many difficulties and challenges for researchers.In this paper,we take the detection and tracking of human targets in infrared thermal video as the research content.The main research contents are as follows:In terms of pedestrian detection,we propose an improved saliency algorithm based on anisotropic filter considering the characteristics of infrared thermal image.Combined with the method of composite threshold segmentation,we can directly obtain candidate regions of human targets in images without using sliding window for multi-scale searching.The multi-feature fusion based on HOG features and CENTRIST features has better local and global contour description capabilities than single features,it fully utilizes the information of human targets in infrared images.Then we use them to train the SVM classifier,our method improve the detection accuracy and reduce false recognition rate.Experiments show that our method has achieved good detection results on the OTCBVS benchmark OSU dataset and our own infrared thermal video data set,and our method can basically achieve real-time effects on the OSU thermal imaging dataset.At the same time,based on the kernelized correlation filters tracking,we establish the description model of the target appearance by using the sparsity-based discriminative classifier and sparsity-based generative model to obtain the likelihood function of the target appearance,then we set a threshold to determine whether the target is occluded,so that we can update the Gaussian response and classifier parameters of the target in the kernelized correlation filters.At the same time,when we establish the appearance model of the target,the adaptive template update method can also ensure that our appearance model can always accurately describe the appearance of target and improve the robustness of tracking as changes of the target appearance during the tracking process.The experimental results show that our method not only can track the target with slight or moderate occlusion in real time,but also can track the target stably after the target being completely occluded for a short time even under certain circumstances,so our method can meet the actual demand.Compared with other algorithms,the experiments show that our improved algorithm can achieve better tracking under different conditions.