Research On Moving Objects Behavior Analysis Based On Correlative Representation Methods

The behavior analysis of moving objects is a difficult problem in the research of unmanned ground vehicles.This article focuses on improving the scene understanding in the complex environment for unmanned ground vehicles,and uses the correlative representation as the entry point to carry out the research on the moving objects behavior analysis method.The main research results and innovations of the thesis are as follows:1.A cascade correlation filtering algorithm based on spatiotemporal saliency is proposed.The algorithm deeply studies the target behavior characteristics in sequential images,uses the correlation filtering to model the target behavior characteristics online,and studies the cascade representation of multi-level features for the boundary effect problem of the correlation filtering algorithm.Use high-level features to express the overall semantics,improving the robustness of associations.Use low-level features to describe local details,improving the accuracy of associations.The introduction of salient semantics effectively alleviates the problem of non-rigid deformation in sequential target associations.The proposed algorithm is compared with the international mainstream algorithms on several international public datasets.The experimental results show that the accuracy and robustness of the algorithm are better than the international mainstream algorithms.2.A background adaptive correlation filtering algorithm based on structural semantic information is proposed.In order to obtain a robust model representation,the correlation expansion algorithm is improved by using a salient expansion matrix,which can reduce the influence of boundary effects while introducing more background information and improving robustness.For the contradiction between semantic information and resolution of deep convolution features,we research structural semantic representation,based on graph theory and non-negative matrix factorization theory.We propose a semantic representation of structural compression,and obtain high-resolution features with structural semantic information.For feature selection,based on structural compressed features,an adaptive feature selection method based on stability factor is proposed,which enables the correlation model to select stable feature representations online and effectively alleviate the occlusion problem of associated targets.Quantitative and qualitative experiments on the international public dataset verify that the proposed algorithm outperforms the international mainstream algorithms in terms of accuracy and robustness.3.A multi-information fusion target association algorithm based on minimum cost flow is proposed.Aiming at the multi-objective local stable association problem,the multi-objective local correlation representation method is studied,and a multiinformation fusion target local correlation expression is established,which can effectively represent the apparent and shape features of the target;Aiming at the global optimization problem for multi-objective association,the multi-objective global optimization model is studied.Based on the target local correlation expression,the multi-objective association problem is modeled by the minimum cost flow method.The proposed algorithm can effectively correlate multiple targets in multiple frames and rank third in the international public database KITTI.4.An algorithm for moving target behavior analysis based on historical frame association is proposed.According to the trajectory characteristics of the moving target,the target behavior category is differentiated,and the target detection and path planning are provided a priori.Based on the target behavior analysis,the fixed-point sensing algorithm is used to enhance the mainstream deep learning-based image detection algorithm.Experiments show that the proposed The algorithm can effectively solve the defects of mainstream algorithms in small target detection.