【摘要】 针对传统手势识别算法准确率不高、鲁棒性不强的问题,基于卷积神经网络提出基于Faster R-CNN的手势识别算法.首先修改Faster R-CNN框架的关键参数,达到同时检测和识别手势的目的;然后提出扰动交叠率算法,避免训练模型的过拟合问题,进一步提高识别准确率.在公共数据集NTU和VIVA上进行手势识别实验的结果表明,该算法有效地避免了训练模型的过拟合问题,比传统算法具有更高的识别准确率和更强的鲁棒性.更多还原

【Abstract】 Due to the problem of low accuracy and robustness of traditional hand gesture recognition, we propose a hand gesture recognition algorithm based on Faster R-CNN, which is evolved from convolutional neural network(CNN). First, we adjust the key parameters of the Faster R-CNN framework to realize the purpose of detecting and recognizing the gestures at the same time. Second, to improve the recognition accuracy of gesture recognition, we propose DisturbIo U algorithm to prevent the network training from over-fitting. We evaluate the improved algorithm on the public NTU-Microsoft-Kinect-hand posture(NTU) dataset and vision for intelligent vehicles and applications(VIVA) dataset. The experimental results show that the proposed algorithm based on Faster R-CNN can effectively avoid the over-fitting problem of the training model, and obtain a better performance in both accuracy and robustness compared with other existing algorithms.更多还原