Research And Implementation Of Improved Pixel Adaptive Segmentation And Tracking Algorithm Based On GPU Acceleration

With the ever-expanding range of video surveillance applications,it is no longer possible to use human-powered surveillance videos.Intelligent monitoring algorithms that automatically process surveillance video are an important development direction in the future.In this paper,we propose an improved pixel adaptive segmentation and tracking algorithm based on GPU acceleration.The algorithm can automatically segment moving objects in the video image and track these targets.The algorithm can be used as the basic algorithm of the intelligent monitoring system.The improved pixel adaptive segmentation and tracking algorithm based on GPU acceleration is essentially a moving target segmentation and tracking algorithm.Therefore,our proposed algorithm can be divided into two algorithm modules: moving target segmentation module and target tracking module.This article will be revolved around our various improvements to these two algorithmic modules.For moving target segmentation(also called foreground detection),the Pixel-Based Adaptive Segmenter(PBAS)method is a very classic algorithm.However,the algorithm has two disadvantages that are difficult to solve.The first drawback is that the algorithm cannot meet higher real-time requirements.In this regard,we use the GPU to process the algorithm's most time-consuming distance feature extraction module to speed up the algorithm.Another disadvantage is that the algorithm is not efficient enough for moving target segmentation in dynamic background scenes or intermittent motion scenes.In this regard,we propose a new moving target segmentation algorithm based on PBAS.We will simply refer to this new algorithm as the We PBAS(The Weight-Pixel-Based Adaptive Segmenter)algorithm.The We PBAS algorithm is superior to the PBAS algorithm in detecting dynamic background scenes and intermittent motion scenes.In the We PBAS algorithm,we first introduce the structure of the weighted background samples.In the background model update phase,unlike the "random update" strategy of the PBAS algorithm,We PBAS uses the "minimum weight update" strategy and the "shortest match distance update" strategy.The "minimum weight update" policy enables the algorithm to replace the least efficient background samples.The "shortest match distance update" strategy allows the algorithm to fine tune the background model to accommodate slow changes in the background.This mechanism improves the detection performace of the algorithm in dynamic background scenes.In addition,we introduce an improved foreground counter that allows the algorithm to adaptively adjust the counter parameters based on the distribution of video distance thresholds to improve detection results in intermittent motion scenarios.On the CDnet2012 dataset,our proposed We PBAS algorithm increased the F-Measure of the PBAS algorithm by 1.55%.On the dynamic background sub-dataset of the CDnet2012 dataset,the proposed We PBAS algorithm is 9.79% higher than the F-Measure of the PBAS algorithm.On the intermittent motion subdataset of the CDnet2012 dataset,our proposed We PBAS algorithm increased the FMeasure of the PBAS algorithm by 3.16%.Since many of the surveillance targets focus on moving targets are only pedestrians,our algorithm is designed to track only the goal of a pedestrian in moving.After the moving target segmentation module detects the bounding box of the moving target,we first use the HOG-SVM(Histogram of Oriented Gradient-Support Vector Machine)pedestrian detection algorithm to determine whether the bounding box is a pedestrian target.After that,we will pass the box information of the pedestrian to the target tracking module.This will effectively prevent the algorithm from tracking the wrong target or foreground noise block.For the detected target tracking problem,the Kernelized Correlation Filters(KCF)method is a prominent target tracking method proposed by the predecessors.However,the KCF method cannot deal with problems such as target size change and partial occlusion.In the target tracking module of this paper,we propose an improved scaleadaptive KCF target tracking method.This method can better deal with the change of target scale through the scale pool strategy.At the same time,we use the foreground region information obtained by the moving target detection module to denoise the input of the tracking algorithm to eliminate the influence of the background on the target tracking algorithm.Compared with the KCF tracking method,our proposed tracking method has better tracking effect on the target when partial occlusion,background clutter,size change,etc.occur.At the end of the paper,we have compared the proposed moving target detection and target tracking methods with the original method qualitatively and quantitatively.From the experimental results,we can see that the proposed algorithm has made significant progress.