Multi-template object tracking method based on CNN and CF

Abstract

The invention discloses a multi-template object tracking method based on CNN and CF. The present invention proposes a method of using a plurality of fixed scale values for normalizing the template before solving the convolution operation, and then reversing the most appropriate scale after obtaining the maximum response value; although the existing DSST algorithm is used to decompose the three-dimensional optimal scale search into the best position in two-dimensional space and the optimal scale mechanism is searched in one-dimensional space, the iteration is slow and the computational complexity is high; in view of the characteristics that the movement of the unmanned aerial vehicle platform is random, the speed is uncertain and so on, the method of the fixed scale value is used to not only meet the tracking algorithm needs, but also to meet the real-time operation; in the characteristic extraction stage, the two characteristics are extracted respectively, two different groups of filters are trained, according to the appearance and the background change of the current object, different weights are set for performing appearance representation of the object; and results obtained from the different characteristics are fused to obtain the tracking result.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and relates to a multi-template target tracking method based on CNN and CF. Background technique [0002] As an important part of computer vision technology, object tracking has a wide range of applications in many fields such as vehicle navigation, human-computer interaction, medical imaging, and video surveillance. The core problem of object tracking is to locate the object in each subsequent frame given the object position in the first frame of the video. The factors affecting target tracking mainly include fast moving target, illumination change, scale change, background interference and occlusion. After continuous research by scholars at home and abroad, the algorithm of target tracking has been developed rapidly. However, with the continuous development of UAV technology, video surveillance has developed from traditional fixed camera surveillance to UAV-based dynamic surveillance....

AI Technical Summary

Problems solved by technology

However, due to the arbitrariness of UAV movement, the relatively large and relatively fast scale and speed changes of the target have caused difficulties for traditional tracking algorithms; (3) Most of the occlusions mentioned in the OTB database are short-term occlusions or Partial occlusion, however, in actual scenes, long-term occlusion often occurs. Due to the lack of corresponding responses in the update mechanism and re-detection mechanism of the existing algorithm, the existing methods do not perform well under the influence of long-term occlusion, and cannot be reproduced. Tracking task failed due to target detection

[0005] In view of the fact that the current tracking algorithm cannot effectively solve the tracking problem of changing viewing angles, it is necessary to design a method to establish multiple templates corresponding to different angles, and to perform multi-angle and multiple target template matching tracking algorithms

Images