Multi-scale target tracking method based on learning rate adjustment and multi-layer convolution feature fusion

Abstract

The invention relates to a multi-scale target tracking method based on learning rate adjustment and multi-layer convolution feature fusion, and belongs to the technical field of computer vision tracking. The method comprises the following steps: firstly, extracting image features by adopting a layered convolutional neural network, and predicting a target position by fusing multi-layer convolutional features by utilizing a linear weighting method; determining the optimal scale of the target by using the target convolution features under multiple scales; and finally, evaluating the confidence coefficient of target response by utilizing average peak value related energy, evaluating the motion condition of the target according to the frame difference mean value and the displacement of two adjacent frames of target images, and adjusting the learning rate of the filter model according to the prediction position confidence coefficient and the appearance change of the target images. Accordingto the method, the traditional correlation filtering tracking method can be effectively processed; traditional manual features lack semantic information, targets and backgrounds cannot be effectivelydistinguished, error accumulation can be caused by indiscriminately updating a filter due to the fact that the learning rate is a fixed value under the conditions of shielding, target loss and the like, and the method can effectively track the targets under complex conditions.

Description

technical field [0001] The invention relates to the technical field of computer vision target tracking, in particular to a multi-scale target tracking method based on learning rate adjustment and fusion of multi-layer convolution features. Background technique [0002] Target tracking is one of the key research directions in computer vision, including machine learning, event detection, signal processing, video surveillance, automatic driving, statistics and other related knowledge. Although target tracking has made great progress in recent years, correlation filter tracking methods still face great challenges due to factors such as occlusion, fast motion, blur, illumination and deformation in tracking targets. [0003] Although the tracking method based on correlation filtering is efficient, the target tracking method based on traditional manual features tends to describe the appearance features of the target and lacks semantic information; in addition, the target will chang...

AI Technical Summary

Problems solved by technology

Although target tracking has made great progress in recent years, correlation filter tracking methods still face great challenges due to factors such as occlusion, fast motion, blur, illumination and deformation in tracking targets.

[0003] Although the tracking method based on correlation filtering is efficient, the target tracking method based on traditional manual features tends to describe the appearance features of the target and lacks semantic information; in addition, the target will change to varying degrees during the tracking process. The learning rate used for template update is a fixed value. In this case, the extracted target samples are used for update, which will cause error accumulation and affect the performance of the classifier.

Images