WAY TO TRAIN DEEP LEARNING NEURAL MODEL USING IMAGES AND DATA FROM POINT CLOUDS AT SAME TIME Russian patent published in 2024 - IPC G01C21/30 G06V20/58 

FIELD: machine learning.

SUBSTANCE: invention relates to the field of machine learning, in particular to image analysis using neural networks. The technical result is to reduce the cost of marking data and improve the accuracy of predicting the surface of the roadway with a neural network model, which will lead to an increase in the accuracy of automatic motion control systems for highly automated wheeled vehicles. The method includes: collecting data in the form of images obtained using cameras and clouds of points obtained using lidars, while manually annotating the clouds of points to highlight the road surface, after which the clouds of points are projected onto camera images using calibration data, then random noise is added in the areas in the image, If the projected points are not covered, the result is a mask that is used as the true position of the road surface in the image, which is necessary to calculate the loss function when training a model where the loss function is a masked version of the cross entropy.

EFFECT: technical result is to reduce the cost of marking data and improve the accuracy of predicting the surface of the roadway with a neural network model, which will lead to an increase in the accuracy of automatic motion control systems for highly automated wheeled vehicles.

1 cl, 6 dwg