A Residual Depth Feature Target Tracking Method Based on Drift Detection

Abstract

The invention discloses a residual depth feature target tracking method for drift detection, which extracts layered features through a convolutional neural network, then adds a residual structure to the convolutional neural network, connects different network layers, and realizes shallow and deep features The fusion of features does not need to be artificially designed. The network structure can automatically realize the function of feature fusion, and use deep features to distinguish objects and backgrounds, which is more discerning than traditional features; when predicting the target position of the current frame, a detection method is proposed. For the strategy of model drift, a response strength drop counter is designed. By comparing the size and count of the response strength of adjacent frames, according to the value of the counter, it is used to judge whether there is model drift, and to take the corresponding model update plan as a remedial measure to achieve Precise tracking.

Description

technical field [0001] The invention relates to the technical fields of image processing and computer vision, in particular to a residual depth feature target tracking method for drift detection. Background technique [0002] In every traffic section, there are surveillance cameras. They observe the passing vehicles all the time, and confirm and track the identity information of the offending vehicles. Among them, the target tracking technology used is already one of the core research topics in the field of computer vision. It has a wide range of applications in real life, not only in traffic monitoring, smart phones, intelligent robots, autonomous driving, military and other fields, it plays an important role. [0003] Traditional target tracking algorithms cannot achieve good tracking results when encountering difficulties such as target deformation, illumination changes, and background clutter, and cannot meet people's needs. With the vigorous development of deep learnin...

AI Technical Summary

Problems solved by technology

In recent years, target tracking has been based on JF Henriques et al. (Henriques J F, Rui C, Martins P, et al. High-Speed ​​Tracking with Kernelized Correlation Filters [J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2014,37(3):583- 596) proposed kernel correlation filtering algorithm is the most representative, has a very high tracking speed, so that people can see the possibility of real-time tracking, but the traditional features used lead to poor tracking effect

In the same year, the deep convolutional neural network proposed by K Simonyan et al. (Simonyan K, Zisserman A.Very deep convolutional networks for large-scale image recognition[J].arXiv preprint arXiv:1409.1556,2014.) was greatly improved in ILSVRC-2014. Brilliant, deep learning began to play a role in computer vision, in 2015 by K He et al (He K, Zhang X, Ren S, et al.Deep Residual Learning for Image Recognition[J].2015:770-778.) The proposed residual network, where the applied residual structure provides a way to train a deeper network, but the real-time performance of tracking cannot meet the requirements well

In subsequent studies, the deep network was used as end-to-end tracking, such as H Nam et al. (Nam H, Han B. Learning Multi-domain Convolutional Neural Networks for Visual Tracking [C] / / Computer Vision and Pattern Recognition.IEEE , 2016:4293-4302.) proposed a multi-domain neural network that takes the original image as input, directly outputs the tracking result, and uses the deep network as a feature extractor, such as M Danelljan et al. (Danelljan M, G, Khan F S, et al.Convolutional Features for CorrelationFilter Based Visual Tracking[C] / / IEEE International Conference on ComputerVision Workshop.IEEE Computer Society,2015:621-629.) The effective convolution operation proposed for tracking, obtained Although we have achieved very good tracking accuracy, we still face the problem that the tracking speed cannot meet the needs of real-time tracking.

At the same time, M Danelljan et al. (Danelljan M, Bhat G, Khan F S, et al. ECO: Efficient Convolution Operators for Tracking [J]. 2016: 6931-6939.) explored the impact of different convolutional layer features on tracking, and concluded that Shallow features are more suitable for the conclusion of tracking; and by C Ma et al. (Ma C, Huang J B, Yang X, et al.Hierarchical Convolutional Features for Visual Tracking[C] / / IEEEInternational Conference on Computer Vision.IEEE Computer Society,2015 :3074-3082.) proposed convolutional layered features for tracking, expounded the impact of neural network shallow and deep features on tracking, effective use of shallow and deep features can significantly improve the tracking effect, but the artificial selection of feature fusion methods Not conducive to accurate target tracking

[0004] The one-sidedness of traditional features and the lack of detection methods and remedies for model drift in traditional tracking models limit the performance of traditional tracking methods

Images