基于深度学习的鲁棒性L1跟踪系统

来源 :第十一届和谐人机环境联合会议 | 被引量 : 0次 | 上传用户：congmingwangzi

【摘要】

：

【作者】

：

高君宇杨小汕张天柱徐常胜

【机构】

：

模式识别国家重点实验室中国科学院自动化研究所,北京 100190

【出处】

：

第十一届和谐人机环境联合会议

【发表日期】

：

2015年7期

【关键词】

：

视频图像 L1跟踪系统深度学习鲁棒性

下载到本地 , 更方便阅读

下载此文赞助VIP

声明 : 本文档内容版权归属内容提供方 , 如果您对本文有版权争议 , 可与客服联系进行内容授权或下架

论文部分内容阅读

基于稀疏表示的L1跟踪方法近年来在视觉跟踪中获得了较大的成功.但其直接使用视额序列各帧内的像素值对模板和目标候选进行表示,而没有考虑到各图像块内部的深度特征信息.发现,在固定摄像头的视频监控场景中,通常可以找到一块区域,该区域中目标物体具有清晰、易于分辨的表观.本文在各视频场景内事先选定一块可以清晰分辨目标表观的参考区域用以构造训练样本,并构建了一个两路对称且权值共享的深度卷积神经网络.该深度网络使得参考区域外目标的输出特征尽可能与参考区域内目标的输出特征相似,以获得参考区域内目标良好表征的特性.最后在L1跟踪框架下使用训练好的深度网培提取目标候选的特征进行稀疏表示,从而获得跟踪过程中应对遮挡,光照变化等问题的鲁棒性.本文在8个行人视频中与当前国际上流行的9种方法对比,结果显示本文提出的方法优于这些已有的跟踪系统.

其他文献

Design Optimization of Direct Heuristic Dynamic Programming Based on Hybrid Estimation of Distributi

As an important class of approximate dynamic programming, the direct heuristic dynamic programming (DHDP) is discussed in this paper.DHDP performs well due to its model-free online learning capability

会议

Approximate Dynamic ProgrammingDirect Heuristic Dynamic ProgrammingEstimation

Link Prediction in Directed Signed Networks Using the Commute Time of Random Walk

In this paper, we studied the problem of link prediction in directed signed social networks.The relationships of these networks can be either positive (friendly) or negative (hostile) and the relation

会议

Directed Signed NetworkLink PredictionCollaborative RecommendationRandom Walk

Stress Test Case Generation of Task Deadlines Using OPL Model

Real-time systems are usually associated with strict performance requirements, especially in safetycritical domains.The bottleneck of performance can be found effectively by stress test.This paper foc

会议

OPL ModelStress Test Case GenerationCOPTask Deadlines

Face recognition algorithm based on Gabor feature and projective dictionary pair learning

In order to improve the recognition rate and speed of face recognition, this paper proposes a face recognition algorithm based on Gabor featutre and projective dictionary pair learning.Firstly, we ext

会议

Face recognitionGabor featureSynthesis dictionaryAnalysis dictionary

Characterizing vertex-degree sequences in scale-free networks

Many large scale complex networks exhibit a scale free vertex degree distribution in a power law form.To better understand the mechanism of power law formation in real world networks, we explore and a

会议

Complex networkScale free networkPower law distributionVertex degree sequence

Disclosure in publishing Chinese information

Published information provides us with the convenience of data analyses and knowledge acquisitions.However it also brings the publishers into the risk of disclosure.All the words, sentences, paragraph

会议

publishing informationdisclosure detectiondecidable problemundecidable proble

Community Evolution Based on Extended Visibility Graph

One of the most important work to analyze networks is community detection.We present a dynamic community discovery method based on Visibility Graph.Firstly, we put forward related definitions of Visib

会议

community evolutionvisibility graphcommunity discoverycomplex network

An unsupervised framework for cytoplasm segmentation from partly superficial-overlapping cervical sm

Cytoplasm segmentation is a subproblem of cervical smear image segmentation.Since the overlapping, complex morphology and other reasons, cytoplasm segmentation is a challenging problem.In this work, w

会议

Cervical smear imageCytoplasm segmentationSLIC superpixelSuperpixel feature e

A Balanced Comparative RNA Secondary Structure Prediction Method Using Weighted Adaboost Extreme Lea

RNA secondary structure prediction is an open problem in bioinformatics.Comparative sequence analysis method is widely used in RNA secondary structure prediction.It takes a group of related RNA molecu

会议

RNA secondary structure predictionComparative sequence analysisCovariation sco

Fast and Accurate Method for Detecting Fingerprint Reference Point

Most methods for detecting fingerprint reference point need to scan the image or orientation field pixel by pixel or block by block to determine whether it is a candidate reference point, which is ine

会议

Reference pointWalkMean-shiftWeighted cross points

基于深度学习的鲁棒性L1跟踪系统

与本文相关的学术论文