计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (9): 151-156.DOI: 10.3778/j.issn.1002-8331.1801-0284

• 图形图像处理 • 上一篇    下一篇

基于卷积神经网络优化TLD运动手势跟踪算法

王  民,李泽洋,王  纯,石新源   

  1. 西安建筑科技大学 信息与控制工程学院,西安 710055
  • 出版日期:2019-05-01 发布日期:2019-04-28

Optimized Motion Gesture Tracking TLD Algorithm Based on Convolution Neural Network

WANG Min, LI Zeyang, WANG Chun, SHI Xinyuan   

  1. School of Information and Control Engineering, Xi’an University of Architecture and Technology, Xi’an 710055, China
  • Online:2019-05-01 Published:2019-04-28

摘要: 针对TLD(Tracking-Learning-Detection)算法在光照变化不均、遮挡严重、跟踪目标模糊等情况下会出现跟踪失败的问题,提出一种基于卷积神经网络优化TLD运动手势跟踪算法。选取手势特征作正样本,其背景作负样本,获取手势HOG特征并投入到卷积神经网络中加以训练,得到手势检测分类器,从而确定目标手势区域,实现手势的自动识别;再利用TLD算法对手势进行跟踪与学习,对正负样本进行估计检测并实时校正,同时运用SURF特征匹配更新跟踪器。实验结果验证,该算法对比TLD经典算法跟踪精度提高了4.24%,增强了运动手势的跟踪效果,相比经典跟踪算法拥有更高鲁棒性。

关键词: 卷积神经网络, TLD算法, 手势跟踪, HOG特征, SURF特征

Abstract: Aiming at the problem that Tracking-Learning-Detection(TLD) algorithm can not track failure under the conditions of uneven illumination variation, occlusion and tracking target ambiguity, an optimized motion gesture tracking TLD algorithm based on convolution neural network is proposed. The gesture is taken as the positive sample and the background is taken as the negative sample. HOG feature is obtained and put into the convolution neural network for training, the gesture detection classifier is gotten. The target gesture area is determined and the automatic recognition of gesture is achieved. TLD algorithm is then used to track and learn the hand gesture, and the positive and negative samples are estimated and detected in real time. At the same time, SURF feature matching is used to update the tracker. Experimental results show that the proposed algorithm can improve the tracking accuracy by 4.24% compared with the original TLD algorithm. This method enhances the tracking effect of motion gesture, which is more robust than the traditional tracking algorithm.

Key words: convolution neural network, Tracking-Learning-Detection(TLD) algorithm, gesture tracking, HOG feature, SURF feature