Dense optical flow estimation system and method based on self-supervised learning

Abstract

The invention provides a dense optical flow estimation system and method based on self-supervised learning, and the dense optical flow estimation system comprises an image pair obtaining unit, an analog estimation unit, a parameter updating unit, and a prediction unit; the image pair acquisition unit is used for extracting a training data set consisting of a large number of original image pairs required for training a to-be-trained optical flow estimation model from any video data, and each image pair comprises a reference image and any frame of target image after the reference image; the parameter updating unit is used for updating convolutional neural network model parameters according to original estimation and amplification estimation results and finally obtaining a convolutional neural network model used for predicting dense optical flow; by adopting the technical scheme of the invention, the speed and the precision of estimating the dense optical flow through the convolutional neural network are high; a self-supervised learning method is adopted, data does not need to be labeled, a large amount of training data can be obtained more easily, and therefore the application capacity of the scheme is improved.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to a dense optical flow estimation system and method based on self-supervised learning. Background technique [0002] In the field of computer vision, dense optical flow (Dense Optical Flow) describes the trajectory of all pixels in an image or the correspondence between a pair of pixels in an image. In image processing tasks such as behavior recognition, target tracking, and motion prediction, optical flow plays a very important role as a motion feature. [0003] In the past imaging tasks, the estimation method of dense optical flow is usually based on deep learning, which has a large amount of calculation and low precision, and cannot be used in practical applications; in addition, it requires a lot of manpower to label all the pixels in the video frame by frame Therefore, supervised learning can only be performed from data synthesized by computer simulation, resulting in serious p...

AI Technical Summary

Problems solved by technology

[0003] In the past imaging tasks, the dense optical flow estimation method is usually based on deep learning, which has a large amount of calculation and low precision, and cannot be used in practical applications; in addition, it requires a lot of manpower to label all the pixels in the video frame by frame Therefore, supervised learning can only be performed from data synthesized by computer simulation, resulting in serious performance degradation in real scenarios

[0004] In summary, the existing dense optical flow estimation methods still have problems such as large amount of calculation, serious dependence on training samples, and poor performance in actual use.

Images