Research On Moving Target Detection Model Based On 3D LARK Feature

With the rapid development of artificial intelligence technology,detecting moving targets has become the hot research topic.The current supervised methods require massive training and the algorithm is very complicated.By contrast,the accuracy of unsupervised method is low and the algorithm relies on the feature.According to the demands of high accuracy,high efficiency and few parameters,this paper explores and designs two models.Aimed at the low contrast ratio,high noise and complex information in untrimmed long videos,this paper firstly presents a space-time local structural statistical matching model based on 3-D weighted locally adaptive regression kernel(LARK).This model mainly contains three parts:composite template set,space-time local matching and overall similarity statistics.Composite template set and test video are matched in spatial-temporal field,then the matching results are statistical counted.Finally the statistical results are analyzed to extract the moving targets,and this model are implemented.Aimed at the various targets with non-compact structure effect the detection accuracy,this paper further presents the double hierarchical structural fusion model based on local structure and neighboring Gaussian structure.The double hierarchical structural fusion model contains four parts:composite template set,neighboring Gaussian structure,space-time statistical matching and double hierarchical structural.Based on the 3-D weighted LARK local structural feature,this paper present the neighboring Gaussian structure,in order to describe the neighboring structural relationship.Then the neighboring similarity and local similarity are evaluated by space-time statistical matching method respectively,which are double hierarchical structural constraints in model.Finally the double hierarchical structure are fused and the actions are extracted.Space-time local structural statistical matching model is compared with Hae Jeo Seo model by detecting moving targets in the same video,the results show that the former can solve the problem of video scene subject to template background.Double hierarchical structural fusion model is compared with S-CNN model by detecting videos in public standard database etc.,the results show that the former is robust to various targets with non-compact structure.Compare with unsupervised methods,the efficient and simple template proposed by our model can achieve equal precision.Meanwhile,the false detection rate can be decreased.