Human body head and shoulder small target tracking method based on improved SiamFC

The invention discloses a human body head and shoulder small target tracking method based on improved SiamFC, and the method comprises the steps: completing the semi-automatic labeling of a data set through a background difference algorithm, and carrying out the manual labeling through LabelImage, and obtaining abundant training samples; introducing an improved channel attention mechanism SE module to improve feature extraction so as to construct an SEMC feature extraction module; an ASPP-M multi-scale cavity convolution module is constructed, an LEDIO module is added based on the ASPP module to carry out preprocessing of multi-scale convolution, and an SURF-PCA module and a Laplace filter are introduced to carry out feature fusion with the multi-scale feature map; a SiamFC backbone network AlexNet feature extraction module is improved, a Gaussian mixture model is introduced for background modeling, an SEMC module and an ASPP-M module are introduced, activation functions Relu and LRN local normalization are i