Moving target detection algorithm based on spatiotemporal correlation multi-channel clustering

Abstract

<sec>In the process of tracking target, certain multi-modal background scenes are not suitable for the off-line training model, and moving target detection is affected as background in the current video environment is mostly multi-modal scene with much noise, and the characters of moving targets irregularly change, which,therefore, requires a more stable and robust moving target detection algorithm. To solve this problem, taking advantage of spatiotemporal relationship learning, the mixed Gaussian model (GMM) is improved in three aspects.</sec><sec>First, the initialization method combining five-frame difference and intra-frame neighborhood average is proposed to obtain the initial parameters of the mixed Gaussian model. The five-frame difference method is introduced to obtain the initial parameters of the model, so that the background model is closer to the real scene. The intra-frame neighborhood average value is introduced, and an accumulation matrix CA is proposed to record the number of neighboring pixel points, then to enhance the information relevant to the neighborhood. This process can reduce the discontinuity of the target. </sec><sec>Second, the calculation method of the neighborhood correlation is introduced to update the parameter of Gaussian model. Since the single pixel feature is related to the neighborhood random correlation, the random subsampling technology and neighborhood spatial propagation theory are combined together, and the execution efficiency is taken into account to simplify the process of updating model. To speed up the model convergence, an observation vector is built in the time dimension to optimize the model parameters, and the weight <i>ω</i> is gained based on the posterior probability. </sec><sec>Then, the color-gradient method incooperated with the color HSI space and gradient information is adopted in this paper to complete the multi-channel Gaussian mixture model. The initial and the updated parameters of the Gaussian model in each channel can be acquired via the above steps. To simplify the computation of three channels, the random sampling of background pixels is introduced. Finally the detection of moving targets in complex environments is realized. </sec><sec>The experiments show that the proposed algorithm has a great improvement in suppressing the influence of complex background and detecting target integrity, and the influence of the moving target in the initial stage is eliminated.</sec>