Human activity recognition (HAR) and cycle analysis, such as gait analysis, have become an integral part of daily lives from gesture recognition to step counting. As the available data and the possible application areas grow, an efficient solution without the need of handcrafted feature extraction is needed. We propose a multi-task recurrent neural network architecture that uses inertial sensor data to both segment and recognise activities and cycles. The solution is validated using three publicly available datasets consisting of more than 120 subjects and 8 activities, 6 of which are cyclic. Our architecture is smaller than comparable HAR models while being robust to different sensor placements and channels. Our proposed solution outperforms or defines state-of-the-art for HAR and cycle analysis using inertial sensors. We achieve an overall activity F1-score of 92.6% and a phase detection F1-score of 98.2%. The gait analysis achieves a mean stride time error of 5.3 $\pm$ 51.9 ms and swing duration error of 0.0 $\pm$ 5.9%. The overall step count error for all activities is −1.5 $\pm$ 2.8%. Thus, we provide a method that is not dependent on feature extraction and a model that is sensor and location independent.

从手势识别到步数计数，人类活动识别（HAR）和周期分析（例如步态分析）已成为日常生活中不可或缺的一部分。随着可用数据和可能的应用领域的增长，需要一种无需手工提取特征的有效解决方案。我们提出了一种多任务递归神经网络体系结构，该体系结构使用惯性传感器数据来分割和识别活动和周期。该解决方案使用三个公开的数据集（包括120多个主题和8个活动，其中6个是循环的）进行了验证。我们的架构比同类的HAR模型要小，但对不同的传感器位置和通道也很稳定。我们提出的解决方案优于或定义了使用惯性传感器进行HAR和循环分析的最新技术。我们实现了92.6％的总体活动F1得分和98.2％的阶段检测F1得分。步态分析得出的平均步幅时间误差为5.3$\pm$ 51.9毫秒，摆幅持续时间误差为0.0 $\pm$5.9％。所有活动的总步数误差为-1.5$\pm$2.8％。因此，我们提供了一种不依赖于特征提取的方法以及一个与传感器和位置无关的模型。