Learning rate progressive updating method based on motion estimation interpolation and dynamic modeling system

Abstract

The invention discloses a learning rate progressive updating method based on motion estimation interpolation and a dynamic modeling system. The method comprises the steps of calculating a motion vector field of a target in real time by using target motion information in video data, using motion estimation interpolation on original video data to generate augmented data frames, and smoothly updatinga model by using the augmented data frames to obtain an accurate prediction model. Through the combination of the prediction model and the learning rate dynamic adjustment module, when the target fast changes, the learning rate is increased, the influence of the quick change target on the model is increased, the adaptive capacity of the model is improved, the learning rate is reduced when shielding interference exists, the interference of the background on the target model is dynamically adjusted and reduced through the learning rate, the robustness of the model is improved, the problem of degradation of the target model in the target dynamic modeling problem is relieved from two aspects, and the target dynamic modeling work is more accurate.

Description

technical field [0001] The invention relates to the field of target dynamic modeling, in particular to a learning rate progressive update method based on motion estimation interpolation and a dynamic modeling system. Background technique [0002] In recent years, in the field of target tracking, many algorithms are closely related to the dynamic modeling of the tracked target. [0003] The main problems they face are: [0004] 1. The model accuracy of the dynamic modeling method with better real-time modeling is insufficient. Recently, characteristic discriminant correlation filters (Histograms of oriented gradients for human detection. Conference on Computer Vision and Pattern Recognition) have been widely used in the tracking field, starting from MOSSE tracker (Visual object tracking using adaptive correlation filters. In: 2010 IEEEComputer Society Conference on Computer Vision and Pattern Recognition.pp.2544-2550). Kernel correlation filter (KCF) (High-speed tracking w...

AI Technical Summary

Problems solved by technology

Kernel correlation filter (KCF) (High-speed tracking with kernelized correlation filters. IEEE Transactions on Pattern Analysis and Machine Intelligence 37(3), 583–596) further studies the improvement of computational efficiency of the loop structure, but the loop structure also leads to Several issues known as boundary effects that affect accuracy

[0005] 2. The real-time performance of the more accurate dynamic modeling method is lacking:

Although this method has good performance in prediction, it has a major disadvantage: specifically, since there is no analytical solution for this optimization objective, the optimization cost of this method is very high

However, most deep learning based methods (Learning multi-domain convolutional neural networks for visual tracking. 2016 IEEE Conference on Computer Vision and Pattern Recognition. pp. 4293–4302) are limited by high computational costs due to the high computational load during training , which makes them unsuitable for real-time object modeling

In addition, the requirement of GPU also increases the cost of equipment, which limits the application of deep methods in real-time embedded systems

[0008] The following conclusions can be drawn from the analysis of related patents at home and abroad: At present, there is no similar application of motion estimation interpolation and progressive update of learning rate to alleviate the problem of target model degradation

Images