Moving target detection method based on time-space domain statistical matching of local features

Abstract

The invention discloses a time-space domain statistical matching and weight allocation combination-based moving target detection method. The method comprises the steps of firstly, proposing and researching the importance of neighborhood pixel points distinguishable for a 3-D LWR operator for extracting finer time-space local features of videos; secondly, performing background removal and multi-scale zooming on features of a 3-D LWR template set to form a composite template set, and performing matching with local features of a test video to obtain a local similarity matrix; and finally, performing time-space statistics to obtain a position probability matrix of a moving target, and extracting the moving target by analyzing the position probability matrix. According to the method, a conventional LARK operator is combined with weight allocation, and a new time-space statistical matching detection model of the moving target is built. Compared with an existing supervision method, the same detection precision can be achieved without a large amount of trainings; and compared with an existing non-supervision method, the scene and the shooting angle suitable for the test video are expanded, the false detection rate is reduced, and the method provided by the invention is suitable for visible light videos and infrared videos.

Description

technical field [0001] The invention belongs to the moving target detection technology in the field of artificial intelligence, in particular to a moving target detection method combining time-space domain similarity judgment, statistical overall similarity and video local weighted features. Background technique [0002] In order to more efficiently extract target information from the ever-increasing mass of videos and improve search efficiency, research on moving target detection models has always been a key development technology in the field of artificial intelligence. Existing supervised methods require a lot of training and have high algorithm complexity, while unsupervised methods have low detection accuracy and depend on features. In this paper, according to the requirements of high detection accuracy, fast timeliness, few parameters and easy implementation, a moving target detection method based on statistical matching of local features in time and space domain is ex...

AI Technical Summary

Problems solved by technology

[0003] The three-dimensional LARK feature was proposed by Seo et al. in 2010. It has rotation and scale invariance, captures the potential structure of the image without being affected by noise, and has the advantages of good stability, but it cannot distinguish the importance of central pixels and neighboring pixels. ; while HOG features have no scale and rotation invariance, LBP features cannot retain image details, and SIFT features are easily affected by background and noise; CNN features extract features of different levels through the convolution kernel from shallow to deep, the effect is good but the training samples are lengthy complex

[0004] In terms of detection methods, Seo’s method in the unsupervised category uses a full-background overall template, and the target matches the template as a whole, resulting in limited applicable scenarios for the video to be tested; When the angle difference is large, when the background is not close to the template, the detection accuracy is very low; the supervised method needs to train the target and the background separately, and then verify and adjust after training, the method process is complicated and inefficient

Images