Target tracking method based on template updating and anchor-frame-free mode

Abstract

The invention discloses a target tracking method based on template updating and an anchor-frame-free mode, which is mainly used for tracking a video target and determining the position and the size of the target in a video. The method comprises the following steps: cutting a training data set; constructing and improving a convolutional neural network; realizing a prediction frame regression structure based on anchor-frame-free and elliptical annotation; updating the target template and carrying out feature fusion to improve the robustness of the template; and using an overlapping ratio regression loss function with distance measurement to improve the convergence effect of training and the fitting degree of the target. According to the method, the problems that in a tracking method, a target template gradually degenerates along with time, and tracking drifts and even is lost are solved, so that the improved network structure is more robust and stable, and a high-precision tracking effect is achieved on the basis of keeping real-time tracking.

Description

technical field [0001] The invention relates to the fields of computer vision, deep learning and image processing, in particular to the fields of feature extraction based on Siamese network, template update and target regression without anchor frame. It specifically relates to a target tracking method based on template updating and anchor-free methods, mainly aimed at the deformation, rotation, occlusion and other changes of the target in the video image during the motion process, which leads to the easy degradation of the algorithm template and the drift and loss of target tracking. And other issues. Background technique [0002] As one of the basic research tasks in the field of computer vision, target tracking is widely used in intelligent monitoring, unmanned driving, security and other fields. In simple terms, target tracking aims to predict the position and shape size of the target in subsequent frames of the video by learning its initial appearance features given the...

AI Technical Summary

Problems solved by technology

However, the preset anchor box needs to set hyperparameters including scale and anchor box aspect ratio. This manual setting will make it difficult for the final tracking box to achieve the optimal fitting effect, thus limiting the upper limit of tracking accuracy.

[0004] In addition, most of the twin network algorithms only use the initial template as the reference frame to search for the target, and do not update the initial target template.

This way of not updating the template will lead to over-dependence on the initial template

When the target is affected by problems such as severe deformation, rotation, occlusion, etc., the feature information will change significantly, which will cause the useful information of the template to decay exponentially with time, and cannot be well matched with the existing target, resulting in target drift or even loss, and occurrence of Issues such as difficulty recovering from tracking failures after drifting

Some algorithms will simply update the samples frame by frame, but the frame by frame update will seriously affect the speed, and the real-time tracking speed cannot be achieved

In terms of loss functions in the algorithm training process, current algorithms use logical loss, Smooth L1 loss, or IOU loss, and the convergence effect of these loss functions on algorithm training is not enough to meet social needs and complex scenarios. When the target is far away, it is difficult to better return to the prediction frame

Images