Moving target tracking method based on multi-target characteristics and improved correlation filter

Abstract

The invention discloses a moving target tracking method based on multi-target characteristics and an improved correlation filter. The method comprises the following steps: inputting position information of a tracked target in a tracking video sequence and an initial frame; extracting multi-channel features of the target to achieve comprehensive information representation of the target; constructing a pixel reliability graph to perform constraint optimization on a correlation filter, and limiting the correlation filter in an image area suitable for tracking; reducing the number of parameters inthe model by using a linear dimension reduction operator, and training a compact sample classification model; performing secondary optimization on the correlation filter through a Gauss-Newton methodand a conjugate gradient method to obtain an optimal correlation filter; responding to the improved correlation filter and the extracted target features of the target search area, and determining theposition of a target tracking box; jointly updating the filter model and the pixel reliability diagram; and outputting a tracking result map. According to the method, moving targets in most scenes can be effectively tracked, and the method has good tracking precision and real-time performance.

Description

technical field [0001] The invention relates to image processing technology, in particular to a moving target tracking method based on multi-target features and an improved correlation filter. Background technique [0002] Vision-based moving target tracking technology takes video images as the processing object, and uses image processing algorithms as the core to track single or multiple moving targets. After decades of research, target tracking algorithms based on computer vision have successively developed many important algorithms at home and abroad, such as optical flow method, Kalman filter, correlation filter, etc., which have achieved better tracking accuracy. Vision-based moving target tracking technology is widely used in both military and civilian fields, such as military strikes, police security and so on. [0003] In the process of moving target tracking, due to reasons such as environment, camera rotation and camera carrier movement, shadow interference, lack ...

AI Technical Summary

Problems solved by technology

[0004] At present, the vision-based target tracking method still has a lot of room for improvement when dealing with the difficulty of target tracking. For example, the patent CN106530330B proposes a video target tracking method based on low-rank sparseness, which uses the statistical tracking method particle filter for state estimation. Low-rank sparse representation of targets and particles is established by establishing a dictionary to reduce computational complexity, and some particles are deleted in advance by reconstruction errors, and then the difference score is calculated to finally determine the target of the next frame. This method requires a sufficient number of samples to be well approximated. The posterior probability density of the system cannot maintain the effectiveness and diversity of particles; the patent CN109828596A proposes a target tracking method, device and UAV, which controls the visible light camera to visually track the target and records the first position of the target in real time. 1. Tracking information, control the infrared camera to track the target in infrared, and record the second tracking information of the target in real time. If the visible light camera loses the target, control the visible light camera to re-lock the target according to the second tracking information and perform visual tracking. Visible light and infrared target features are not fused, alternate tracking does not update historical target information, and it is easy to cause tracking loss; patent CN109729498A proposes a target tracking method and system based on Voronoi diagram for adaptive node selection, using Voronoi diagram As a network model, the nodes in the network are clustered. There are active nodes, dormant nodes and the only cluster head node in the cluster. According to the proposed node selection algorithm, the sensor nodes in t

Images