Ellipse detection and positioning in complex vision measurement environments

Ellipses are often widely used as a visual positioning marker. Occluded ellipses in complex backgrounds are difficult to detect and locate accurately. Traditional ellipse detection methods are prone to missed detection and large edge positioning deviations. This greatly affects the accuracy of subsequent algorithms. A single-stage anchor-free ellipse detector is proposed in this paper, which achieves fast and high-precision detection and positioning of ellipses in complex occluded environments. According to the special shape of the ellipse, the attention mechanism is introduced to help the model more accurately locate the area containing the ellipse and pay more attention to useful channel information. Secondly, dynamic upsampling and variable convolution are used to design the decoder of the model to improve the positioning accuracy of the ellipse. Gauss-Wasserstein is used for IoU loss, and the center point deviation loss is modified to optimize the ellipse regression accuracy. Combined with real measurement applications, this paper builds an ellipsoidal sphere dataset for large-scale spatial camera calibration. Experiments on self-built datasets and public datasets show that the performance of the proposed method is better than the state-of-the-art model.