Moving target detection algorithm is to detect the changing region in the input image sequence, so as to extract the target from the background. It is very important for subsequent target recognition and tracking. Due to the complex and changeable natural environment, there is a large number of dynamic backgrounds and changing lights in complex weather, such as rain, snow, fog, vegetation shaking and water surface fluctuation, which has always been the primary problem of moving target detection in complex scenes. In this paper, when the camera has a fixed angle of view, the background modeling algorithm is adopted. To suppress the problems of dynamic background, slow target absorption and image coding noise, based on ensuring real-time performance, the time-frequency domain and frequency-domain characteristics of the input image are analyzed. Because the dynamic background fluctuates in a certain gray range. Through the texture features processed by LBP, the influence of illumination change can be suppressed. An adaptive threshold moving target detection algorithm based on texture features is proposed. First of all, the algorithm converts the input image sequence into a grayscale image, and uses the perceptual hashing algorithm to calculate the average pixel value in the window of 3×3 to remove the high-frequency details in the image part and improve the calculation speed. Then, the dynamic background and noise are processed preliminarily. The frequency domain of the input image is analyzed, and the maximum frequency of the gray frequency distribution histogram of each pixel is counted as the reference background and compared with each frame to obtain the difference matrix. The standard deviation of all values greater than 10 and less than 100 in the difference matrix is calculated as the adaptive threshold, and the gray value of each frame is corrected. Then the dynamic background and noise are processed a second time. Firstly, the Local Binary Pattern (LBP) is used to process the reference background and each frame of the image after preliminary processing to obtain the LBP value. Then, the hamming distance 3 was selected as the threshold to correct the LBP value of each frame. Finally, the LBP texture feature map of each frame is analyzed from the frequency domain, and the background modeling and foreground extraction are carried out according to the polymorphic frequency distribution. To suppress the influence of illumination change, dynamic background and noise on foreground extraction, this paper proposes an improved Hash_LBP algorithm combined with a perceptual hash algorithm and uses Hamming distance to constrain it. Experiments show that the proposed algorithm can effectively suppress noise, illumination change and dynamic background in a variety of complex scenes such as infrared and visible light, quickly and accurately extract foreground targets, and the algorithm is also effective in dynamic background suppression for ViBe and GMM algorithms.