Target tracking method based on coding and decoding structure

Abstract

The invention discloses a target tracking method based on an encoding and decoding structure. According to the method, a similar generative adversarial network structure is generated through the combination of an encoder-decoder and a discriminator, the features extracted by an encoder are more generalized, and the essential features of a tracked object are learned. Due to the fact that the objects which are semi-shielded and affected by illumination and motion blur exist in the object frames, the influence on the network is smaller, and the robustness is higher. According to the method, FocalLoss is used for replacing a traditional cross entropy loss function, so that the loss of easy-to-classify samples in the network is reduced, the model pays more attention to difficult and misclassified samples, and meanwhile the number of positive and negative samples is balanced. Distance-U loss is used as regression loss, an overlapping region is concerned, other non-overlapping regions are concerned, scale invariance is achieved, the moving direction can be provided for a bounding box, and meanwhile the convergence speed is high.

Description

technical field [0001] The invention belongs to the field of image processing and computer vision, and in particular relates to an object tracking method based on an encoder-decoder structure. Background technique [0002] One of the main goals of computer vision is to enable computers to replicate basic functions of human vision, such as motion perception and scene understanding. To achieve the goal of intelligent motion perception, much effort has been devoted to visual object tracking, which is one of the most important and challenging research topics in computer vision. Essentially, the core of visual object tracking is to reliably estimate the motion state (i.e., position, orientation, size, etc.) of the target object in each frame of the input image sequence. At this stage, the target tracking algorithm mainly has two major branches, one is based on the correlation filtering algorithm, and the other is based on the deep learning algorithm. The target tracking method ...

AI Technical Summary

Problems solved by technology

The L1 loss function regresses the four coordinate values ​​separately, but the loss function has the following disadvantages: first, when using the loss function to calculate the bounding box regression loss of the target detection, the loss of the four points is independently calculated, and then added to obtain The final bounding box regression loss, the assumption of this approach is that the four points are independent of each other, but in fact the four points are correlated; secondly, the actual evaluation frame detection index is to use the intersection ratio, which is not equivalent to L1

However, the pre-training task based on the ImageNet image database is based on the classification task, and the features extracted by the convolutional neural network are also more suitable for the classification task, which does not match the feature extraction requirements for the tracking task.

Images