Target tracking method based on novel response graph fusion

Abstract

The invention discloses a target tracking method based on novel response graph fusion, and the method comprises the steps: taking the position of a given first frame of target object as the center, cutting out a search region block, generating a corresponding training mark graph, carrying out the training of a depth regression model, enabling a preset loss function to be minimum, and obtaining a trained depth regression model; taking the position of the target object predicted by the previous frame as the center, cutting out search area blocks with the same size, inputting the search area blocks into the trained depth regression model, generating a final response graph through feature extraction and response graph fusion, and enabling the maximum value of the response graph to represent the position of the predicted target object; after the position of the target object is obtained, performing scale estimation on the size of the target object; and updating the depth regression model according to the search area block of the historical frame and the corresponding training mark graph, and repeating the above steps to predict the position and size of the target in each subsequent frame. According to the invention, the target position can be positioned more accurately, so that the target tracking accuracy is improved.

Description

technical field [0001] The invention relates to a target tracking method. In particular, it concerns an object tracking method based on novel response map fusion. Background technique [0002] Object tracking is a promising but difficult field of research in computer vision algorithms that has won widespread acclaim for its wide range of applications in autonomous driving, traffic flow monitoring, surveillance and security, human-computer interaction, and medical diagnostic systems. Object tracking is an active research area in computer vision due to opportunities and different tracking challenges. In the past decades, many research teams have made great efforts, but object tracking still has great potential to be further explored. The difficulty of object tracking lies in countless challenges, such as occlusion, background clutter, illumination changes, scale changes, low resolution, fast motion, loss of sight, motion blur, deformation. [0003] Discriminative correlatio...

AI Technical Summary

Problems solved by technology

[0005] However, in the current target tracking method based on the depth regression model, the traditional response map fusion method used is a simple addition or multiplication fusion method. Complex challenging situations are not robust and can easily cause bounding box drift problems

Images