Fall Recognition Based on Multi-feature Learning Fusion Cascade Classification
-
摘要: 为解决不同跌倒方式的训练误差无法充分收敛而导致分类效果欠佳的问题，在提取穿戴式设备加速度、角速度及合加速度三维特征的基础上，提出了一种级联分类的跌倒识别(Cascade Fall Recognition，CFR)方案；同时为降低级联串行运算的复杂度，进一步提出了基于经验特征维度降维的低复杂度级联分类跌倒识别(Low-complexity Cascade Fall Recognition，LCFR)方案。最后，面向SisFall数据集，分别使用单个多分类器方案、CFR方案和LCFR方案进行四分类任务(日常行为，向前跌倒，横向跌倒，向后跌倒)，对F1-Score和训练时间复杂度进行对比。实验结果表明：CFR方案在四分类跌倒任务时的F1-Score达98.56%；在接近无损F1-Score的前提下，LCFR方案的训练时间复杂度比CFR方案降低了61.11%，且该方案的F1-Score高于同类跌倒识别方案。Abstract: To solve the problem that the training errors of different fall methods cannot converge sufficiently, which leads to poor classification effect, a Cascade Fall Recognition (CFR) scheme for cascade classification is proposed based on the extraction of three-dimensional features of acceleration, angular velocity and combined acceleration of wearable devices. At the same time, to reduce the complexity of Cascade serial operation, a low-complexity Cascade Fall Recognition (LCFR) scheme based on dimensionality reduction of empirical features is further proposed. Finally, for SisFall dataset, a single multi-classifier scheme, CFR scheme and LCFR scheme were used to perform four classifications (daily behavior, forward fall, lateral fall and backward fall), and the F1-Score and training time complexity were compared. The experimental results show that the F1-Score of CFR scheme is 98.56% in the four-category fall task classification. On the premise of approaching lossless F1-score, the training time complexity of LCFR scheme is 61.11% lower than that of CFR scheme, and the F1-score of scheme is higher than that of similar fall recognition schemes.
-
Key words:
- fall recognition /
- cascade classification /
- feature dimension reduction
-
图 5 常规的跌倒识别方案输出的混淆矩阵图
Figure 5. Confusion matrix diagram of the output of conventional fall recognition scheme
表 1 融合后的跌倒类型
Table 1. Types of falls after fusion
跌倒行为 跌倒类型 F1、F4、F5、F6、F8、F10、F13 向前跌倒 F2、F11、F14 向后跌倒 F3、F7、F9、F12、F15 横向跌倒 
下载: 导出CSV
表 2 窗口特征量提取表
Table 2. The feature extraction in a window
特征量 数学表示 维度 加速度最大值 max ai (i=x, y, z, sum) 1×4 加速度最小值 min ai (i=x, y, z, sum) 1×4 加速度平均值 ${\bar a_i} = \frac{1}{N}\sum\limits_{k = 1}^N {{a_i}} [k]\quad (i = x, y, z, {\rm{sum}})$ 1×4 加速度标准差 ${\sigma _{{a_ - }i}} = \sqrt {\frac{1}{N}\sum\limits_{k = 1}^N {{{\left( {{a_i}[k] - {{\bar a}_i}} \right)}^2}} } \quad (i = x, y, z, {\rm{ sum }})$ 1×4 加速度峰度 ${K_{a\_i}} = \frac{{\frac{1}{N}\sum\limits_{k = 1}^N {{{\left( {{a_i}[k] - {{\bar a}_i}} \right)}^4}} }}{{{\sigma _{a\_i}}}} - 3\quad (i = x, y, z)$ 1×3 加速度偏度 ${S_{{a_ - }i}} = \frac{{\frac{1}{N}\sum\limits_{k = 1}^N {{{\left( {{a_i}[k] - {{\bar a}_i}} \right)}^3}} }}{{\sigma _{{a_ - }i}^3}}\quad (i = x, y, z)$ 1×3 加速度最大差值 $\Delta {a_i} = \max {a_i} - \min {a_i}\quad (i = x, y, z)$ 1×3 加速度最大差值斜率 ${k_{a\_i}} = \frac{{\max {a_i} - \min {a_i}}}{{{t_{\max }} - {t_{\min }}}}\quad (i = x, y, z, {\rm{sum}})$ 1×4 合加速度累计差值 ${{\mathop{\rm cum}\nolimits} _{\Delta {a_ - }{\rm{sum }}}} = \sum\limits_{k = 1}^{N - 1} {\left( {{a_{{\rm{sum }}}}[k + 1] - {a_{{\rm{sum }}}}[k]} \right)} $ 1×1 合加速度积分 $f[k] = \int {{a_{{\rm{sum }}}}} {\rm{d}}t$ 1×1 角速度最大值 $\max {\omega _i}\quad (i = x, y, z)$ 1×3 角速度标准差 ${\sigma _{\omega \_i}} = \sqrt {\frac{1}{N}\sum\limits_{k = 1}^N {{{\left( {{\omega _i}[k] - {{\bar \omega }_i}} \right)}^2}} } \quad (i = x, y, z)$ 1×3 角度最大差值 $\Delta {\varphi _i} = \max {\varphi _i} - \min {\varphi _i}\quad (i = x, y, z)$ 1×3 角度最大差值斜率 ${k_{{\varphi _ - }i}} = \frac{{\max {\varphi _i} - \min {\varphi _i}}}{{{t_{\max }} - {t_{\min }}}}\quad (i = x, y, z)$ 1×3 注：${a_{{\rm{sum}}}} = \sqrt {a_x^2 + a_y^2 + a_z^2} $；ai[k]、ωi[k]分别表示滑窗内第k个采集点的加速度、角速度；${\bar \omega _i} = \frac{1}{N}\sum\limits_{k = 1}^N {{\omega _i}} [k]\quad (i = x, y, z)$；${\varphi _i} = \int {{\omega _i}} {\rm{d}}t(i = x, y, z)$；N为在2.5 s滑窗内采集的数据个数。
下载: 导出CSV
表 3 常规的跌倒识别方案测试结果
Table 3. Test results of conventional fall recognition scheme
% 行为 F1-Score Precision Recall 日常行为 99.85 99.94 99.75 向前跌倒 93.43 90.32 96.76 向后跌倒 96.76 96.04 97.49 横向跌倒 89.05 92.15 85.83
下载: 导出CSV
表 4 三层级分类器的准确率
Table 4. The accuracy of three-level classifier
% 层级 SVM KNN RF 第一级 99.31 99.60 99.67 第二级 97.67 97.77 98.45 第三级 89.96 89.48 95.04
下载: 导出CSV
表 5 CFR方案的测试结果
Table 5. Test results of CFR Scheme
% 行为 F1-Score Precision Recall 日常行为 99.89 98.80 99.98 向前跌倒 98.58 99.38 97.78 向后跌倒 98.75 100.00 97.52 横向跌倒 97.02 96.45 97.60
下载: 导出CSV
表 6 各层级分类器的训练数据集及目标分类
Table 6. The training data set and target classification for each level of classifier
层级 分类任务 样本类型 目标分类 第一级 二分类 日常行为与跌倒行为 日常行为与跌倒行为 四分类 日常行为与跌倒行为 日常行为、向后跌倒、向前跌倒、横向跌倒 第二级 二分类 跌倒行为 向后跌倒与其他跌倒(向前跌倒、横向跌倒) 四分类 跌倒行为 向后跌倒、向前跌倒、横向跌倒
下载: 导出CSV
表 7 第一级分类器不同累计特征贡献度的准确率与训练时长
Table 7. The accuracy and training time for different cumulative feature contributions of the first level classifier
累计特征贡献度/% F1-Score/% 训练时长/ms 60 99.59 1 486 70 99.62 2 337 80 99.67 2 446 100 99.76 5 663
下载: 导出CSV
表 8 3个方案的分类器训练时长与F1-Score
Table 8. The training time and F1-Score of three schemes
方案 F1-Score/% 训练时长/ms 常规的跌倒识别方案 94.77 6 684 CFR 98.56 6 835 LCFR 98.48 2 658
下载: 导出CSV
表 9 跌倒识别研究结果对比
Table 9. The comparison of other research results for fall recognition
方案 分类任务 F1-Score/% SVM方案 三分类 92.46 DT方案 三分类 82.53 RFC方案 三分类 90.77 XGBoost方案 三分类 90.16 LCFR方案 四分类 98.48
下载: 导出CSV
-
[1] KALACHE A, FU D, YOSHIDA S, et al. World health organisation global report on falls prevention in older age[R]. Geneva: World Health Organization, 2007. [2] COSTA J E, ANDRADE R M D C, ROCHA L S, et al. Computational solutions for human falls classification[J]. IEEE Access, 2021, 9: 161590-161602. doi: 10.1109/ACCESS.2021.3132796 [3] 徐甲栋, 陈强, 徐一雄, 等. 基于MEMS传感器的实时跌倒检测系统设计[J]. 传感器与微系统, 2022, 41(7): 77-80. https://www.cnki.com.cn/Article/CJFDTOTAL-CGQJ202207020.htmXU J D, CHEN Q, XU Y X, et al. Design of real time fall detection system based on MEMS sensor[J]. Transducer and Microsystem Technologies, 2022, 41(7): 77-80. https://www.cnki.com.cn/Article/CJFDTOTAL-CGQJ202207020.htm [4] 赵珍珍, 董彦如, 曹慧, 等. 老年人跌倒检测算法的研究现状[J]. 计算机工程与应用, 2022, 58(5): 50-65. https://www.cnki.com.cn/Article/CJFDTOTAL-JSGG202205005.htmZHAO Z Z, DONG Y R, CAO H, et al. Research status of elderly fall detection algorithms[J]. Computer Enginee-ring and Applications, 2022, 58(5): 50-65. https://www.cnki.com.cn/Article/CJFDTOTAL-JSGG202205005.htm [5] ANDO B, BAGLIO S, LOMBARDO C O, et al. A multisensor data-fusion approach for ADL and fall classification[J]. IEEE Transactions on Instrumentation & Mea-surement, 2016, 65(9): 1960-1967. [6] YU M, GONG L, KOLLIAS S. Computer vision based fall detection by a convolutional neural network[C]//Proceedings of the 19th ACM International Conference on Multimodal Interaction. New York: ACM, 2017. [7] SALEH M, JEANNES R L B. Elderly fall detection using wearable sensors: a low cost highly accurate algorithm[J]. IEEE Sensors Journal, 2019, 19(8): 3156-3164. doi: 10.1109/JSEN.2019.2891128 [8] ZURBUCHEN N, WILDE A, BRUEGGER P. A machine learning multi-class approach for fall detection systems based on wearable sensors with a study on sampling rates selection[J]. Sensors, 2021, 21(3): 938/1-23. doi: 10.3390/s21030938 [9] KWON S B, PARK J H, KWON C, et al. An energy-efficient algorithm for classification of fall types using a wearable sensor[J]. IEEE Access, 2019, 7: 31321-31329. doi: 10.1109/ACCESS.2019.2902718 [10] NANCY G, PANKAJ D K. An argumentation enabled decision making approach for fall activity recognition in social IoT based ambient assisted living systems[J]. Future Generation Computer Systems, 2021, 122: 82-97. doi: 10.1016/j.future.2021.04.005 [11] HARRIS A, TRUE H, HU Z, et al. Fall recognition using wearable technologies and machine learning algorithms[C]//2016 IEEE International Conference on Big Data. Washington: IEEE, 2016. [12] 裴利然, 姜萍萍, 颜国正. 基于支持向量机的跌倒检测算法研究[J]. 光学精密工程, 2017, 25(1): 182-188. https://www.cnki.com.cn/Article/CJFDTOTAL-GXJM201701024.htmPEI L R, JIANG P P, YAN G Z. Research on fall detection system based on support vector machine[J]. Optics and Precision Engineering, 2017, 25(1): 182-188. https://www.cnki.com.cn/Article/CJFDTOTAL-GXJM201701024.htm [13] HUSSAIN F, EHATISHAM-UL-HAQ M, AZAM M A, et al. Activity-aware fall detection and recognition based on wearable sensors[J]. IEEE Sensors Journal, 2019, 19(22): 4528-4536. [14] SYED A S, KUMAR A, SIERRA-SOSA D, et al. Determining fall direction and severity using SVMs[C]//2020 IEEE International Symposium on Signal Processing and Information Technology. Louisville, Kentucky: IEEE, 2020. [15] ANGELA S, LÓPEZ J, VARGAS-BONILLA J. SisFall: a fall and movement dataset[J]. Sensors, 2017, 17(12): 198-207. doi: 10.3390/s17010198 [16] PUTRA I, VESILO R. Window-size impact on detection rate of wearable-sensor-based fall detection using supervised machine learning[C]//2017 IEEE Life Sciences Conference. Sydney: IEEE, 2017. [17] 宁梓涵, 高熳祺, 陈志遥, 等. 后仰跌倒人体撞击加速度及冲量分析[J]. 医用生物力学, 2018, 33(6): 551-557. https://www.cnki.com.cn/Article/CJFDTOTAL-YISX201806014.htmNING Z H, GAO M Q, CHEN Z Y, et al. Analysis on impact acceleration and impulse during backward falling[J]. Journal of Medical Biomechanics, 2018, 33(6): 551-557. https://www.cnki.com.cn/Article/CJFDTOTAL-YISX201806014.htm [18] 丁蕊, 汤庸, 曾伟铨, 等. 基于分类算法的潜在好友推荐系统[J]. 华南师范大学学报(自然科学版), 2017, 49(6): 124-128. doi: 10.6054/j.jscnun.2017169DING R, TANG Y, ZENG W Q, et al. A potential friend recommendation system based on classification algorithm[J]. Journal of South China Normal University(Natural Science Edition), 2017, 49(6): 124-128. doi: 10.6054/j.jscnun.2017169 [19] 何坚, 周明我, 王晓懿. 基于卡尔曼滤波与k-NN算法的可穿戴跌倒检测技术研究[J]. 电子与信息学报, 2017, 39(11): 2627-2634. https://www.cnki.com.cn/Article/CJFDTOTAL-DZYX201711013.htmHE J, ZHOU M W, WANG X Y. Wearable method for fall detection based on Kalman filter and k-NN algorithm[J]. Journal of Electronics & Information Technology, 2017, 39(11): 2627-2634. https://www.cnki.com.cn/Article/CJFDTOTAL-DZYX201711013.htm -
点击查看大图
图(8) / 表(9)
计量
- 文章访问数: 24
- HTML全文浏览量: 6
- PDF下载量: 1
- 被引次数: 0
