深度学习在人体行为识别中的比较研究

需积分: 9 4 下载量 121 浏览量 更新于2024-09-08 1 收藏 2.06MB PDF 举报
“面向人体行为识别的深度特征学习方法比较” 本文深入探讨了人体行为识别问题,特别是基于智能手机惯性加速度传感器数据的深度特征学习方法。研究中对比了两种主流的深度学习技术——深度卷积神经网络(Deep Convolutional Neural Networks, DCNNs)和长短期记忆网络(Long Short-Term Memory, LSTM)。这两种方法都在解决序列数据分类问题上具有显著能力,尤其适用于处理传感器数据的时间序列特性。 首先,传感器数据通过重叠加窗的预处理步骤,以提取关键信息并减少噪声。预处理是数据科学中的关键步骤,能够提高模型对原始数据的理解和处理能力。 接着,处理后的带标签样本数据直接输入深度网络模型。端到端学习是深度学习的一个重要特点,它允许模型直接从原始数据学习,而无需人工设计特征。这种自动化特征工程的方法减少了人力投入,并且往往能挖掘出更深层次的特征。 在实验部分,研究者使用了UCI机器学习知识库中的人体行为识别数据集。这个公开数据集广泛用于验证和比较行为识别算法的效果。实验结果显示,应用Dropout技术的深度卷积神经网络在行为识别任务中的准确率达到了90.98%。Dropout是一种防止过拟合的策略,通过在训练过程中随机忽略一部分神经元,可以增加模型的泛化能力。 相比之下,长短期记忆网络LSTM专门设计用于处理序列数据,擅长捕捉时间依赖性。然而,本研究中DCNNs在行为识别任务上的表现优于LSTM,这可能是因为DCNNs在处理图像和空间模式方面更为有效,尽管惯性传感器数据本质上是时间序列,但其空间模式(如身体部位的运动)也可能被有效地捕获。 结论指出,Dropout深度卷积神经网络是一种高效的深度特征学习方法,对于人体行为识别有显著的效果。这项研究为基于智能手机的实时行为监控提供了理论基础和技术支持,对于健康监测、安全监控等领域有着广泛应用前景。同时,也强调了在选择深度学习模型时应考虑数据的特性以及不同模型的适用场景。 关键词:深度学习,行为识别,序列数据分类,深度卷积神经网络,长短期时间记忆网络 中图分类号:TP391 文献标志码:A 文章编号:1001-3695(2018)09-2815-03 doi:10.3969/j.issn.1001-3695.2018.09.059
2018-04-13 上传
Current developments in nanotechnology make electromagnetic communication possible at the nanoscale for applications involving Body Sensor Networks (BSNs). This specialized branch of Wireless Sensor Networks, drawing attention from diverse fields such as engineering, medicine, biology, physics and computer science, has emerged as an important research area contributing to medical treatment, social welfare, and sports. The concept is based on the interaction of integrated nanoscale machines by means of wireless communications. One key hurdle for advancing nanocommunications is the lack of an apposite networking protocol to address the upcoming needs of the nanonetworks. Recently, some key challenges have been identified, such as nanonodes with extreme energy constraints, limited computational capabilities, Terahertz frequency bands with limited transmission range, etc., in designing protocols for wireless nanosensor networks. This work proposes an improved performance scheme of nanocommunication over Terahertz bands for wireless BSNs making it suitable for smart e-health applications. The scheme contains – a new energy-efficient forwarding routine for electromagnetic communication in wireless nanonetworks consisting of hybrid clusters with centralized scheduling; a model designed for channel behavior taking into account the aggregated impact of molecular absorption, spreading loss, and shadowing; and an energy model for energy harvesting and consumption. The outage probability is derived for both single and multilinks and extended to determine the outage capacity. The outage probability for a multilink is derived using a cooperative fusion technique at a predefined fusion node. Simulated using a Nano-Sim simulator, performance of the proposed model has been evaluated for energy efficiency, outage capacity, and outage probability. The results demonstrate the efficiency of the proposed scheme through maximized energy utilization in both single and multihop communication; multisensor fusion at the fusion node enhances the link quality of the transmission.