A Visual Tracking Method Based on Convolutional Sparse Filtering

Abstract

The invention discloses a visual tracking method based on convolution sparse filtering, comprising the following steps: 1) using the convolution sparse filtering method for offline training, and performing unsupervised feature learning on the tracking video sequence to obtain a set of convolution kernels 2) Convolutional neural network is formed with this convolution kernel to perform online tracking, so as to extract features from the input image; 3) Combine the kernelized correlation filter tracking framework to achieve target tracking. The invented visual tracking method based on convolution sparse filtering, which is based on the principle of deep learning, automatically learns the characteristics of the target to be tracked, so that the advantages of big data can be used to obtain more stable and discriminative features, thereby achieving high precision , High robust tracking effect. The invention is based on the convolution sparse filtering tracking method, and has the characteristics of high accuracy, fast speed and stable tracking effect.

Description

technical field [0001] The invention belongs to the technical field of visual tracking, in particular to a visual tracking method based on convolution sparse filtering. Background technique [0002] Visual tracking is a hot spot in the field of computer vision research and has a wide range of applications. The tracking technology has high requirements on the computing speed, and the inability to achieve real-time means that it is difficult to be practical. At present, the best tracking methods are based on deep learning. Among them, the best tracking algorithm is MDNet proposed by Hyeonseob Nam and Bohyung Han of Pohang University of Technology in South Korea. This method is based on the multi-domain learning framework of convolutional neural network (CNN), which separates domain-independent information from specific domain information. to get a valid representation. For the first time, it is possible to directly use video sequences as training data. In addition, they in...

AI Technical Summary

Problems solved by technology

They reported that the algorithm can reach 100fps, but these measures reduce the tracking performance of the algorithm. According to the tracking effect they reported, the tracking performance is similar to the traditional DSST, SAMF, KCF algorithms

Moreover, the reported speed is obtained on the GPU platform. For example, the algorithm can reach 165fps on the Titan X GPU, and 100fps on the GTX 680GPU, but it can only reach 2.7fps on the CPU, which is also difficult to be practical.

From the analysis of the existing algorithms, it can be seen that the method based on deep learning uses tracking video sequences for training, and obtains an effective representation of the data, thereby obtaining a better tracking effect, but the deep network depth brings large calculations. It is difficult to meet the effect of real-time tracking, which makes it difficult to apply such algorithms in engineering practice

Images