Method for video moving object detection based on relative statistical characteristics of image pixels

A method for video dynamic object detection based on relative statistical characteristics of image pixels, characterized in that:

• step 1:
• pixel classification stage:
• in the training set with a plurality of images denoted as T, extracting a gray value of each pixel to form a vector as basic information of each pixel value and performing clustering processing, and dividing pixels in the images into several categories based on brightness characteristics, each category generating a cluster center as a representative in the category;
• an expression of pixels in a pixel space being
• S={I1,I2,I3,I4. . . IT}
• S={(u,v,t)|0uW,0vH,0tT};
• wherein, I1, I2, I3, I4 . . . IT are an image frame sequence, (u, v, t) are pixel points in the pixel space, W, H, T are respectively a frame width, a frame height, and a number of frame of the image frame sequence;
• step 2:
• processing complex pixels stage:
• calculating a Euclidean distance for each pixel to its cluster center, and extracting the pixel points that are more different from the cluster center as complex points for further processing, and then re-processing the remaining pixels for classification, a formula for expressing the Euclidean distance of the pixel being
• d
• =
• 1
• T
• ⁢
• ∑
• i
• =
• 1
• T
• (
• p
• i
• -
• q
• i
• )
• 2
• wherein, pi is a brightness value of a target pixel at the i-th frame, qi is a brightness value of a cluster center point at the i-th frame;
• step 3:
• reference point selection stage:
• for each target point, adopting a method of image spatial clustering, and selecting the reference points scattered in each category;
• for the complex points, according to their changing trends, finding two categories that best represent it, and selecting the reference points in the two categories respectively;
• step 4:
• detection stage:
• processing and comparing the target pixel to be detected with its reference point, setting a threshold to determine whether it is a foreground point;
• wherein:
• an object of comparison in the step 4 is whether a difference between gray values of the pixel to be detected and its corresponding reference point meets requirements;
• a specific processing method is as follows:
• for general static background points, a gray value of each pixel point in a training set and its reference point are directly calculated and averaged to obtain an inherent gray value difference;
• for complex dynamic background points, because it uses two classes are collectively represented, each class corresponds to a specific number of frames in the training set, that is, a number of effective frames;
• difference of reference points of each class in a sequence of corresponding effective frames is calculated to obtain its own inherent gray difference;
• wherein the model and parameters are set and stored in a training stage, without a need to update a background model and parameters in real time;
• in the detection stage, an input video is converted to image mode and adjusted to a specified size;
• respectively modulating a coordinate of a reference point of each pixel, a difference between a gray value of a reference point and a target point;
• for complex dynamic background points, respectively modulating the reference point coordinates of a decomposition class, an effective frame number subscript of a corresponding reference point of each class, and a difference between a gray value at an effective reference frame number and that at the target point;
• for each pixel point, if it is a general point, that is, a static background point, obtain a difference after subtracting its gray value from a gray value of the pixel of a reference point, and subtracting an inherent gray value, if it is within an adequate threshold, it is determined that this pixel is a point in a background of the image, and there is no foreground object passing here, otherwise it is determined to be a foreground point;
• for complex points, that is, dynamic background points, same determination operation is performed on two types of reference points to which they belong;
• setting a threshold nq, if after each reference point is used for determination, a number of times that it is determined as a previous scenic spot exceeds the threshold nq, then this pixel is finally output as a foreground pixel.