A Face Tracking Method Based on Regression Algorithm and Its Application

Abstract

The invention discloses a face tracking method based on a regression algorithm and application. According to the method, the prior information that the to-be-tracked target is the human face is fullyutilized, so that the tracking algorithm is more targeted, and the result of the tracking frame can be more accurate. The designed face tracking algorithm is mainly composed of a face tracking networkand a group of strategies and can simultaneously complete face frame tracking and judge whether a tracking result is still a face or not. According to the designed face tracking network, a single andsmall neural network is adopted, and end-to-end training and rapid prediction are achieved. The input of the designed face tracking network is only related to the current frame image information andthe face frame position of the previous frame, and is not related to the face images of the previous frame and the previous frame. In addition, a tracking algorithm and a face detection algorithm arecombined to form a complete face detection tracking algorithm, and the real-time detection tracking application of the face in the video can be realized.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and relates to a face tracking method and application based on a regression algorithm. Background technique [0002] In applications such as embedded video face recognition (such as face attendance machines, etc.), the application of face tracking means is of great significance for real-time video processing; the accuracy of tracking results can also further reduce the subsequent analysis modules (such as human face quality judgment). [0003] Existing tracking methods mainly include methods based on correlation filtering (such as KCF, etc.), methods based on convolutional neural networks, and other methods. These methods have the following problems: [0004] 1. It does not use the prior information that the target to be tracked is a human face, but only tracks a specified (or pre-framed) visual target (maybe a human eye, a mouth, a cat, etc. ). [0005] 2. When the target is rotated ...

AI Technical Summary

Problems solved by technology

[0004] 1. It does not use the prior information that the target to be tracked is a human face, but only tracks a specified (or pre-framed) visual target (maybe a human eye, a mouth, a cat, etc. )

[0005] 2. When the target is rotated or deformed, it is often difficult for existing tracking methods to obtain an accurate tracking frame (for example, when the face changes from the front to the side, etc.)

[0006] 3. For the disappearance of the tracking target, the occlusion of the tracking target, the back of the tracked face target, etc., the general tracking algorithm (such as KCF) can only give the most likely position (tracking frame) of the tracking target in the image, but cannot It is very good to give a very effective judgment of whether the tracking frame is no longer a target to be tracked (especially when the process of such a target disappearing, occluding, and turning to the back of the head is slow)

[0007] 4. For single-scale target tracking methods such as KCF, when the scale of the target changes, it is difficult to obtain an accurate tracking frame. Adding multiple scales on the basis of these algorithms is difficult to include all scales, and it will also bring a rapid increase in time-consuming

On the whole, TLD has a time-consuming problem

Images