A Convolutional Neural Network Based Adaptive Feature Selection Target Tracking Method

Abstract

A method for video target tracking based on adaptive feature selection for convolutional neural networks, comprising the following steps: 1) multi-layer CNN feature extraction; 2) training correlation filters; 3) feature selection; 4) the target tracking process is as follows: using conv3 ‑2, conv3‑4, conv3‑8, conv3‑12, conv3‑16 The five-layer convolution feature is used as the target feature observation, and five correlation filters w are trained after feature selection k , k∈[1, 5], and then finally determine the target position (x * ,y * ), the weight parameter w k It is updated online according to the learning rate ρ. The invention significantly reduces the feature dimension, and reduces the amount of calculation without losing the feature discrimination ability.

Description

technical field [0001] The invention belongs to the technical field of video target tracking, in particular to an adaptive feature selection target tracking method. Background technique [0002] Target tracking is the process of feature extraction, appearance modeling, motion analysis and target association of the target of interest in the image sequence. It is widely used in intelligent transportation, visual servoing, human-computer interaction and other fields. Although the field of target tracking has made great progress in recent years, it still faces great challenges, such as target occlusion, illumination changes, pose changes, complex background, low resolution and fast movement of targets, etc., which will cause target drift or even tracking failure. [0003] With the great success of convolutional neural network in target detection and object recognition, more and more researchers have begun to focus on applying convolutional neural network features to the field of...

AI Technical Summary

Problems solved by technology

[0004] Aiming at the problems of feature dimension disaster, feature redundancy, and low robustness caused by the existing convolutional neural network features, the present invention proposes a feature based on the distance between convolutional neural network feature maps. Selection method; this method incorporates multiple correlation filter weights to determine the target position at the same time; uses information propagation strategy to adaptively select features

Images