Journal of Frontiers of Computer Science and Technology ›› 2020, Vol. 14 ›› Issue (5): 841-847.DOI: 10.3778/j.issn.1673-9418.1906063

Previous Articles     Next Articles

Application of Convolutional Neural Network in Dynamic Gesture Tracking

LI Dongjie, LI Dongge, YANG Liu   

  1. College of Automation, Harbin University of Science and Technology, Harbin 150080, China
  • Online:2020-05-01 Published:2020-05-08

卷积神经网络在动态手势跟踪中的应用

李东洁李东阁杨柳   

  1. 哈尔滨理工大学 自动化学院,哈尔滨 150080

Abstract:

In order to solve the tracking of dynamic gesture targets in complicated scenarios, an improved YOLOv3 (you only look once) real-time tracking algorithm is put forward for dynamic gestures. First, concerning the problem of poor real-timeliness of YOLOv3 on-line inspection, the main network structure of YOLOv3 is improved by using the characteristics of single target inspection of gestures. Second, an inspection tracking method suitable to the planned area of gesture tracking under complicated scenarios is put forward to inspect gesture targets, screen impacts of targets not being tracked at the moment in the background and complete real-time tracking of gestures. Last, training and testing are conducted in a unified manner to designed gesture data. Experimental results show that, for gesture tracking under complicated scenarios, the algorithm outperforms YOLOv3 algorithm and related target tracking algorithms.

Key words: dynamic gestures, YOLO (you only look once), real-time tracking, convolutional neural network (CNN)

摘要:

为了解决复杂场景中对动态手势目标进行跟踪的问题,提出了一种动态手势的改进YOLOv3实时跟踪算法。算法首先针对YOLOv3网络检测实时性较差的问题,利用对手势这样的单类目标检测的特性对YOLOv3的主干网络结构进行改进。其次提出一种适合于复杂场景下手势跟踪的规划区域的检测跟踪方法,对手势目标进行检测,过滤掉背景中非当前跟踪目标造成的影响,完成对手势的实时跟踪。最后在设计的手势数据集中进行训练和测试。实验结果表明,算法在复杂场景中的手势跟踪性能均优于YOLOv3算法和一些相关目标跟踪算法。

关键词: 动态手势, YOLO, 实时跟踪, 卷积神经网络(CNN)