基于深度学习的人体行为识别研究

Human action recognition based on deep learning

为解决传统人体行为识别算法存在的运动前景检测不准确、特征提取模糊以及训练识别耗时长等问题,本文提出了基于深度学习的人体行为识别研究方法。利用骨架提取方法对运动前景进行检测及特征提取;针对人体行为动作的时序性,提出了连续帧组合方法;在模型训练环节,对比了不同的网络模型参数,选择了最优的激活函数、优化算法以及dropout系数。最后,结合网络模型,分类识别测试样本集中的各种行为,并将识别的结果和当前流行的算法进行比较,通过对比实验,最终实验结果证明了本文所提方法优于其他方法,平均识别率相比其他方法有较大的提高。

In order to solve the problems of inaccurate motion foreground detection,feature extraction blur and training recognition time-consuming in traditional human behavior recognition algorithm,a deep learning-based human behavior recognition research method is proposed.The skeleton extraction method is used to detect the motion foreground and feature extraction;for the timing of human behaviors,a continuous frame combination method is proposed;in the model training process,different network model parameters are compared and the optimal activation function,optimization algorithm and dropout coefficient are selected.Finally,combined with the network model,this study classifies and identifies various behaviors in the test sample set and compares the recognition results with the current popular algorithms.Through comparative experiments,the experimental results show that the proposed method is superior to other methods and there is a big improvement in the average recognition rate.