A moving target tracking method and device

Abstract

The invention provides a moving target tracking method and apparatus. The method specifically includes the steps of, during a feature extraction phase, extracting features based on directional vectors and a maximal average pooling method, establishing the mapping relation between the directional vectors and initial sampling sheets, and compensating an initial sampling sheets set; and during a movement estimating phase, using reconstruction errors to correct the observation probability of the sampling sheets, increasing the weights of the sampling sheets corresponding to the directional vectors, and calculating corrected observation probability to obtain the final moving target tracking result. The method and apparatus effectively compensate insufficient of sampling information during an initialization process, characteristic types are enriched, and more comprehensive expression of movement target is realized. The accuracy and stability of the tracking result are improved.

Description

technical field [0001] The invention relates to the field of moving target tracking, in particular to a moving target tracking method and device. Background technique [0002] Visual tracking algorithms based on sparse representation usually combine sparse representation and motion estimation methods to track moving objects, such as combining sparse representation with learning strategy, mean shift method, Bayesian motion estimation, particle filter, etc. to track moving objects in the environment. [0003] Using sparse optical flow to extract the features required in the tracking process can achieve the purpose of online tracking, but this method is only suitable for small displacement moving objects, and cannot accurately track displacement information larger than the local image structure. In addition, the particle filter method cannot avoid the drift phenomenon when the same obstacle appears. Although the robustness of the algorithm can be improved by learning each par...

AI Technical Summary

Problems solved by technology

[0003] Sparse optical flow is used to extract the features required in the tracking process. Although the purpose of online tracking can be achieved, this method is only suitable for moving objects with small displacements, and cannot accurately track displacement information larger than the local image structure.

In addition, the particle filter method cannot avoid the drift phenomenon when the same obstacle appears.

Although the robustness of the algorithm can be improved by learning each particle as a unique work, the speed of the algorithm will be greatly reduced, and the problem of particle degradation still exists

The Bayesian motion estimation method cannot track moving objects online, and cannot achieve the purpose of online real-time tracking

Images