Research On Moving Foreground Extraction And Tracking For Intelligent Video Surveillance

With the development of technology,the intelligent video surveillance technology has gradually become a research hotspot in the field of computer vision.It can automatically analyze the video sequence acquired by the monitoring device to realize the extraction of moving foreground,the tracking and recognition of the target in the scene.The technology of motion foreground extraction and tracking which is the key technology for intelligent video surveillance is still one of the research difficulties in the field of academia and industry.Based on the related research of domestic and foreign,this thesis makes a research on the moving foreground extraction and target tracking in the complex environments.The main work are as follows.(1)This thesis analyzes the classic foreground extraction algorithm-ViBe which constructs background model based on pixel sampling,then proposes a new foreground extraction algorithm.Firstly the proposed algorithm constructs an approximate real background model,then uses the spatial continuity of pixel movement to compare the pixel with the neighborhood pixels which are in the background model in order to complete the segmentation of the foreground and background.And during the background model update process it sets up a "protective band" area to prevent the misjudged background from eroding the foreground.The proposed algorithm is compared with some state-of-the-art methods on the benchmark.Experimental results show that the proposed algorithm has lower rate of misjudged pixel and can achieve excellent segmentation precision.It also has favorably robustness to the dynamic background and camera jitter.(2)For the kernelized correlation filters target tracking algorithm cannot well deal with the change of target scale and posture,this thesis proposes a scale adaptive tracking algorithm based on the original algorithm.The proposed algorithm samples image patches at multiple scales,and selects the image patch with the highest response value at each scale as the candidate sample.Then these candidate samples are passed through a positive sample library in order to select the appropriate sample as final tracking result.The proposed algorithm is compared with some state-of-the-art methods on numerous benchmark sequences.Experimental results show that the proposed algorithm can effectively track the target with scale and posture change,and it has favorably performance on robustness.(3)This thesis combines the motion foreground extraction module and tracking module into a whole framework.After the extraction algorithm obtains moving foreground,the tracking algorithm begins tracking the target.The framework is tested on numerous video sequences.Experimental results show that the framework can extract the moving foreground effectively even if the camera is jittered,and in the tracking process it can effectively adapt to changes in target scale,posture.