Target matching method for discriminant cascade appearance model based on multi-feature fusion

Abstract

A target matching method for a discriminant cascade appearance model based on multi-feature fusion is provided. The method comprises the following steps: carrying out online acquisition of a training sample set based on the MIL, and using a sample to carry out feature training and learning; extracting multiple feature fusion for a target template to form a robustness feature descriptor, determining the weight of the descriptor through multiple experiment combinations, and using the improved B's distance to measure the similarity; using the Top-rank index to carry out matching performance evaluation on the RGB histogram, the HOG and other fusion features; based on the Adaboost algorithm, carrying out continuous learning and training, and carrying out theoretical derivation and experimental verification to make the error rate of the classifier to be descended exponentially until the convergence is stable; taking the training sample data set X={Xi| Xi is an element of a set {X+, X }} as input to carry out training, and obtaining a discriminant appearance model H (x) of the cross-view sample combination; and updating the tag of the sample set online for online learning, and using the obtained example probability to re-weight the priori estimate result of the particle filter.

Description

technical field [0001] The invention belongs to the technical field of image processing, and relates to a target matching method of a discriminative cascade appearance model based on multi-feature fusion. Background technique [0002] Visual object tracking is a hot issue in the field of machine vision research. Its main task is to obtain the trajectory of the object of interest in the video sequence. It has been widely used in video surveillance, intelligent transportation and other research fields. At present, challenging problems still need to be solved: 1) internal factors such as pose changes, scale changes, and deformation of the tracking target itself; 2) external factors such as brightness changes, partial occlusions, and complex backgrounds in the tracking process. [0003] In recent years, experts have focused on combining video tracking research in overlapping areas with machine learning theory. The main research in the early stage is to propose a tracking algorit...

AI Technical Summary

Problems solved by technology

However, based on this model, there are still feature redundancy in the process of feature extraction, which violates the principle of fusion feature complementarity, and a large amount of feature storage and calculation are required in the model update, and the number of classifiers cannot be determined. In addition, the current discriminant Tracking is susceptible to problems such as noise interference

On the whole, since video tracking often involves the process of feature extraction, fusion, model building and updating, etc., the defects in the process will not only limit the efficiency and other aspects, but also affect the reliability of the model due to the environmental noise in the surrounding area. , the adaptability and expansion performance of the trained model and parameters are also poor, which will lead to drift and misjudgment in complex situations such as illumination changes and pose changes.

Images