Font Size: a A A

A Study On Visual Tracking Technology For Large Field Of View And Small Target

Posted on:2020-01-05Degree:MasterType:Thesis
Country:ChinaCandidate:C Y LiuFull Text:PDF
GTID:2428330599959248Subject:Mechanical engineering
Abstract/Summary:PDF Full Text Request
Visual tracking technology is often used in video surveillance,virtual reality,sports events,and so on.Limited by algorithms and platforms,fast tracking with large field of view and small targets is hard to achieve.At present,visual detection and tracking algorithms still couldn't reach the accuracy and robustness in human level,and the image acquisition in visual tracking system is hard to be achieved with a high frame rate,high resolution and large field of view simultaneously.This paper studies the problems existing in the visual detection,tracking algorithm and tracking platform in the practical application,and detailed researches are as follows:(1)A novel Multi-Scale Single Shot Multibox Detector is proposed,as single shot detectors have poor performance on targets at different scales and small targets.The proposed algorithm re-calibrates the importance of the channel and executes the weighted fusion operation on the channel through the attention fusion network;then generates the feature map suitable for the target detection under different scales through the multi-scale fusion network.Compared with the original algorithm on the VOC07 Benchmark,the proposed algorithm accuracy increases by 1%,which can handle similar targets at different scales and small targets very well.(2)A novel Attention-Based Multi-Layer Correlation Filters Tracking Algorithm is proposed,as traditional correlation filters tracking algorithms have poor performance on targets at fast motion and background clutters.The proposed algorithm uses the large margin spatial-temporal regularized correlation filters as sources,and generates multi-layer model in different spatial and temporal dimensions;in addition,a mean-and-noise based metric function is proposed to evaluate the importance of each model.Compared with the other 8 tracking algorithms on the OTB50 Benchmark,the proposed algorithm has the highest accuracy and success rate,and the improvement is more obvious in the case of fast motion and background clutters,and the speed of 25 fps also meets real-time demands.(3)A Visual Tracking System Based on Saccade Mirror is developed,as the traditional tracking systems can't consider the high frame rate,high resolution and large field of view simultaneously.The system uses a convex lens group to enlarge the camera's field of view,and uses dual camera structure to track the target,which can achieve high frame rate,high resolution and large field of view simultaneously,and the filed of view is 1.7~9 times larger than before.Finally,a visual tracking software is developed for the visual tracking system.The software invokes the above-mentioned visual detection and tracking algorithms to perform the task of target detection and tracking.Through a large number of experiments,the effectiveness of the visual tracking system is proved.After the image acquisition and control components are added,the frame rate of the system is 20 fps.
Keywords/Search Tags:Visual Detection, Visual Tracking, Correlation Filter, Attention Mechanism, Multi-Layer Model, Saccade Mirror, Tracking System
PDF Full Text Request
Related items