To address the issues of low image quality of birds, difficulty in accurately testing the velocity and trajectory of birds flocks, and low accuracy of test results in traditional birds ingestion experiments, a method for measuring key parameters of aero-engine birds ingestion tests based on deep learning was proposed. Firstly, an image segmentation network based on U-Net was built to automatically segment the contour of the birds, achieving precise positioning of the edge area of the birds to be detected. Then, a key frame merging algorithm was developed, which used similar pixels in the images before and after the segmented flying object area to realize automatic annotation of feature points of the birds in continuous frame images. Finally, based on Kalman filtering and cubic spline interpolation methods, an automatic fitting model for the movement trajectory of the birds was developed to achieve high-precision measurement of the flight trajectory and velocity. The experimental results show that, compared with manual precise calibration, the centroid coordinates obtained by the designed centroid discrimination algorithm have an average error of less than 1.00%; compared with manual measurement, the velocity obtained by the velocity measurement algorithm has an error of less than 2.00%.
HOU et al. (Sun,) studied this question.