A target tracking method and system for a rotor-operated flying robot

Abstract

The invention belongs to the technical field of visual target tracking, and discloses a method and system for target tracking of a rotor-operated flying robot. Based on the Siamfc framework, through offset learning, Resnet50 is introduced as a feature extraction network, so that the network can learn more semantic information , to cope with changes in the appearance of the target; the tracking network adds a target scale estimation module on the basis of the classification discriminator, which can predict the IOU between the target bounding box and the target real frame, accurately predict the target bounding box, and use the reverse gradient to The bounding box is iteratively corrected so that the network can accurately predict the scale change of the target; the output of Resnet50 multi-layer features is used, and the strategy of residual fusion is adopted to fuse the outputs of different layers of the network to further improve the robustness of the algorithm. Improve the performance of the network, and ensure the ability of the network to distinguish small targets, and finally achieve accurate tracking of the target.

Description

technical field [0001] The invention belongs to the technical field of visual target tracking, and in particular relates to a target tracking method and system for a rotor-operating flying robot. Background technique [0002] At present, visual target tracking is an important research direction in computer vision, and has a wide range of applications, such as: video surveillance, human-computer interaction, unmanned driving, etc. In the past 20 to 30 years, the visual target tracking technology has made great progress, especially in the last two years, the target tracking method using deep learning has achieved satisfactory results, making the target tracking technology a breakthrough. [0003] Target tracking technology has a very rich application in the field of UAVs. The automatic detection and tracking technology system of battlefield targets has become the basis for UAVs to realize situational awareness and precise strikes on the battlefield. The airborne computer furt...

AI Technical Summary

Problems solved by technology

[0006] The difficulty in solving the above problems and defects is: by introducing a deep network such as ResNet50, in theory, the feature extraction network can learn more abundant target information. This kind of performance will degrade the performance of the tracking algorithm

In addition, the prediction of the target scale module is a complex task, which requires the network to learn more semantic information rather than some shallow feature information. Therefore, a deeper feature extraction network is also required, and an effective target scale estimation module is also designed. Many factors need to be considered, such as how to embed into the tracking network, the design of the receptive field of the network, and how to fuse features. A scale estimation module that does not match the target discrimination and classification module will lead to a decrease in performance

Images