In order to improve the accuracy of the eye tracking algorithm and ensure a certain image processing speed, this paper proposes to combine the deformable convolution method to improve the feature distribution extraction level. The fixed-size sampling in the standard convolution makes it difficult for the learning network to adapt to the geometric deformation of the image. In order to solve this limitation, the deformation modeling ability of deformable convolution is used to add a certain offset variable to the position of each sampling point in the convolution kernel. So as to achieve the extraction of potential features, the single frame of the original image is described. According to the current research, the deformable convolution has made preliminary applications in the field of computer vision. After comparing with the advanced eyeball positioning tracking detection network experiment, the accuracy of the deformable convolutional YOLO network can reach 0.685, and the average image processing speed can reach 42 frames per second, which is better than the original YOLO network and the advanced eyeball and location tracking detection network.