基于多域卷积神经网络跟踪的动态手势识别

Dynamic gesture recognition based on multi-domain convolutional neural network

针对传统的手势跟踪方式在复杂环境下跟踪效果差导致动态手势识别准确率不高的问题,提出了一种多域卷积神经网络跟踪框架下的动态手势识别算法。该算法采用多域卷积神经网络实现手势跟踪环节,并根据手势跟踪的特点,精简网络结构,构建全域通用fc6层,增强网络对动态手势跟踪的适用性,提升跟踪效果。其次,采用VGG-19对跟踪网络构建的动态手势轨迹特征图谱进行识别。算法将跟踪问题简化成目标与背景的二分类,采用多域卷积神经网络学习跟踪目标共性,能更好地给出跟踪目标模型,且浅层卷积神经网络的利用更能强化空间信息,从而提高动态手势跟踪和识别的效果。通过两组数据库测试表明,对自建的动态手势库识别率高达97.5%,并在Chalearn Gesture Data国际标准手势数据库取得了93.33%的识别率,验证了算法的有效性。

In view of the traditional gesture tracking method,the poor tracking effect limits the accuracy of dynamic gesture recognition.A dynamic gesture recognition algorithm for the multi-domain convolutional neural network tracking framework is proposed.First,multi-domain convolutional neural network is used to implement gesture tracking in the algorithm,and according to the characteristics of gesture tracking,the network structure is streamlined,and the universal fc6 layer is constructed.The applicability of dynamic gesture tracking is enhanced while the tracking effect is improved in this algorithm.Second,the VGG-19 network is used to identify the dynamic gesture trajectory feature map constructed by the tracking network.The tracking problem is reduced to a binary classification of target and background,and multi-domain convolutional neural networks are used to learn commonality of tracking targets.The shallow convolutional neural network is used to enhance the spatial information.Thus,the dynamic gesture tracking and recognition result is improved.Based on the Chalearn Gesture Data International Standard gesture database,The recognition rate of the four types of self-built dynamic gestures is as high as 97.5%,and 93.33%recognition rate which demonstrates the effectiveness of the algorithm.