Research And Application Of Background Subtraction Algorithm In Video Surveillance Scene

Moving object detection technology is widely used in video surveillance scenes,which is the basis of subsequent higher-level visual analysis tasks,and the most in-depth research in moving object detection technology is the background subtraction algorithm.Background subtraction algorithm is a key step of many computer vision application,because it can detect the changed target in the video stream without any prior knowledge of the scene.Background subtraction algorithm has developed for nearly 30 years,and it has developed into a new research field.In this emerging field,researchers have made remarkable achievements by continuously introducing research methods,theoretical achievements and innovative technologies from other disciplines.In practical application,background subtraction needs to overcome many challenges in the scene to play an effective role in complex environment.The research goal of this thesis is to improve the performance of the classical background subtraction algorithm and propose a new background subtraction algorithm for the practical application of video surveillance scenes.Based on this goal,this thesis first reviewed the development course of background subtraction method and studied various algorithms.,using objective performance evaluation metrics and the general CDnet 2012 datasets to study the advantages and disadvantages and adaptability of the algorithm,and then optimized the PBAS algorithm with better performance,and improved algorithms are proposed from various angles,finally,for the high-resolution small-target projectile falling point detection,a heterogeneous multi-scale background subtraction algorithm(HMBS)is proposed,this algorithm is suitable for heterogeneous computing platforms represented by Zynq.The main innovations of this thesis are:1.An improved optimization algorithm for distance calculation in the PBAS is proposed.The ratio of the gradient amplitude to the pixel value in the foreground segmentation process is measured by calculating the average value of the gradient amplitude of the image.2.An improved algorithm for updating the distance threshold in the PBAS is proposed.Different update methods are used for the front points and background points,and the recall of the algorithm in this area is improved by reducing the distance threshold of the front points in the dynamic background area.3.For the scene of camera jitter,this thesis proposed a new non-parametric model based on PBAS algorithm to measure the dynamic degree of the background in the scene of camera jitter.4.In the actual application scenario of projectile falling point detection,an HMBS algorithm based on heterogeneous calculation is proposed.The HMBS algorithm introduces multi-scale calculations to coordinate the use of computing resources of different architectures.In foreground detection,the neighborhood subtraction method is used to deal with camera jitter interference,and the area level threshold is used to deal with image noise interference.Based on the background subtraction,a multi-frame confirmation process is introduced to further strengthen the correctness of the algorithm's detection results and ensure the accuracy of the Bomb-fall target.In the simulation experiment based on the CDnet 2012 dataset,compared with the original PBAS algorithm,the PBAS algorithm proposed in this thesis has improved the F1 value of all directory scenarios,and the overall F1 value has increased by 4.18%.Among them,the F1 value in the Dynamic Background directory has been improved by 15.6%,indicating that the improved PBAS algorithm has played a significant optimization effect in scenes with dynamic backgrounds.In the optimization algorithm for camera jitter,based on the Camera Jitter directory for comparison experiments,the F1 value is increased by 15.4%compared with the original PBAS algorithm,and the misclassified pixels are reduced by 0.51%.It can be seen that the improved algorithm also reduces the number of missed detections while reducing the number of false detections.In the simulation experiment of projectile falling point detection,HMBS can correctly detect all 7 projectile falling point targets in the test video of more than 60000 frames without any false detection,after the HMBS algorithm was transplanted to the Zynq platform,the real-time requirements of the detection system were met,the results show that the algorithm design meets the expectation.