Video Moving Objects Detection And Paralle Implementation With Constrained Energy Optimization

Video moving object detection,as an important part of intelligent video surveillance system,is the basis of a series of subsequent processing,such as object analysis,identification,tracking and so on.With the widespread use of intelligent video surveillance system in social security,intelligent traffic management,industrial production control and other fields,the application scenarios become more diverse.This put forward higher requirements on performance of video moving object detection in accuracy,stability,adaptability,etc.Graph cuts algorithm based on energy minimization has a significant effect on image segmentation,but produces boundary bias in continuous video moving object detection.On the basis of the existing research,this thesis optimizes the energy function of the graph cuts based moving object detection algorithm by energy constraint,and the new algorithm is parallel accelerated to meet the requirements of real-time processing.The specific research work are as follows:The moving object detection algorithm based on graph cuts would produce boundary bias in continuous video detection.Through further analysis,we have found that,lacking of effective manual interaction during continuous video detection makes the boundary term of energy function fixed,which brings the problem.In response,this thesis proposes a video moving object detection algorithm based on gradient weighting with boundary length constraint.The algorithm crystallizes the intrinsic link between the moving object length and image gradient value.By extracting the length characteristics of the moving object as prior information to constrain the gradient,this algorithm realize the adaptive optimization of boundary term,which reduce the boundary bias.The experimental results show that,compared with other similar methods,the detection results of the proposed algorithm are smoother,and have stronger adaptability in different scenarios.The proposed video moving object detection algorithm has improved the object detection accuracy,but also greatly increases the computational complexity simultaneously.Based on the CUDA platform,this thesis implements a high efficient parallel moving object detection algorithm by three aspects: realizing coarse-grained and fine-grained parallelism by parallel task partitioning,improving data throughput by optimizing GPU memory allocation,enhancing instruction level parallelism by eliminating the instruction branch.The experimental results show that the GPU parallel implementation of the proposed algorithm significantly increases the computation speed compared with the CPU serial implementation.The parallel implementation can achieve real-time detection as 30 fps for 432×288 resolution video,and 12 fps for 768×576 resolution,whichimproves the real-time performance of the moving object detection algorithm.