·386· 智能系统学报 第13卷 在关联关系是难以用现有的方法准确表达出来的， 2013:1-4 通过以上方法提高跌倒检测的准确性。在分析跌倒 [8]Duan K B,Keerthi SS.Which is the best multiclass SVM 检测要求的基础上，针对RNN的特点，设计了序列 method?An empirical study[M]//Multiple Classifier Sys- 化的传感器数据、RNN训练与检测输入表示方法， tems.Springer Berlin Heidelberg,2005:278-285 给出了用于跌倒检测的RNN训练算法以及基于 [9]BREIMAN L.Random forest[J].Machine learning,2001, RNN的跌倒检测算法，将跌倒检测转换为输入序列 45(1):5-32. [10]SZ Erdogan,TT Bilgin,J Cho.Fall-down detection by us- 的分类问题；并在本文前期所实现的基于分布式神 ing K-nearest neighbor algorithm on WSN datalC]// 经元大规模RNN系统的基础上，在Spark平台上实 GLOBECOM Workshops.Houston,USA,2011: 现了基于RNN的跌倒检测系统，使用Fall adl data 2054-2058. 数据集进行了检测效果和对接近跌倒数据检测能力 [11]WU G.XUE S.Portable preimpact fall-down detector with 的测试与分析，验证了RNNFD能有效提高跌倒检 inertial sensors[J].IEEE Transactions on Neural Systems 测的准确率，F值相比现有跌倒检测系统能提高 Rehabilitation Engineering,2008,16(2):178-183. 12%和7%，并且能够检测出接近跌倒行为。 [12]OJETOLA O,GAURA E,BRUSEY J.Data set for fall- 下一步将进一步优化用于跌倒检测RNN down events and daily activities from inertial sensors 的训练方法，同时考虑使用LSTM减少时间开销，提 [C/ACM Multimedia Systems Conference.Oregon,USA, 2015:243-248. 高准确率。 [13]CHEN J,KWONG K,CHANG D,et al.Wearable sensors 参考文献： for reliable fall-down detection[Cl//International Confer- ence of the IEEE,NEW YORK,USA,2006:3551-3554. [1]吴天吴.基于3轴加速度传感器及陀螺仪的老年人摔倒 [14]LI Q,STANKOVIC J A,HANSON M A,et al.Accurate 识别D1.北京：北京工业大学，2013：2-3 fast fall-down detection using gyroscopes and acceleromet- WU Tianhao.Identification of old people's fall-downing er-derived posture information[C]//Sixth International based on 3-axis acceleration sensor and gyroscope[D]. Workshop on Wearable and Implantable Body Sensor Net- Beijing,China:Beijing University of Technology,2013: works.CA,USA,2009:138-143. 2-3. [15]VAIDEHI V,GANAPATHY K,MOHAN K,et al.Video [2]ROUGIER C,MEUNIER J,ST-ARNAUD A.3D head based automatic fall-down detection in indoor environ- tracking for fall-down detection using a single calibrated ment[Cl//International Conference on Recent Trends in In- camera[J].Image and Vision Computing,2013,31(3): formation Technology.Chennai,INDIA,2011:1016- 246-254 1020. [3]MUBASHIR M,SHAO L,SEED L.A survey on fall-down [16]BOSCH-JORGE M,SANCHEZ-SALMERON A J, detection:Principles and approaches[J].Neurocomputing, ANGEL VALERA,et al.Fall-down detection based on the 2013,100(2):144152, gravity vector using a wide-angle camera[J].Expert Sys- [4]MATHIE M J,BASILAKIS J,CELLER B G.A system for tems with Applications,2014,41(17):7980-7986. [1刀佟丽娜，宋全军，葛运建.基于时序分析的人体摔倒预测 monitoring posture and physical activity using acceleromet- 方法).模式识别与人工智能，2012,25(2少：273-279 ers[Cl/Interational Conference of the IEEE Engineering in TONG Linna,SONG Quanjun,GE Yunjian.Time series Medicine and Biology Society.Istanbul,Turkey,2001: analysis based human fall-down prediction method[J].Pat- 3654-3657 tern Recognition&Artificial Intelligence,2012,25(2): [5]卢先领，王洪斌，王莹莹，等.一种基于加速度传感器的人 273-279. 体跌倒识别方法[].计算机应用研究，2013,30(4)： [18]GIBSON R M,AMIRA A,CASASECA-DE-LA- 1109-1111 HIGUERA P,et al.An efficient user-customisable mul- LU Xianling,WANG Hongbin,WANG Yingying,et al.Hu- tiresolution classifier fall-down detection and diagnostic man fall-downing detection based on accelerometer[J]. system[C]//International Conference on Microelectronics. Computer Application Research,2013,30(4):1109-1111. Changchun,China,2015:228-231. [6]FANG-YIE LEU,CHIA-YIN KO,YI-CHEN LIN,et al. [19]LUO D.LUO H.SCHOOL I.Fall-down detection al- Smart Sensors Networks[M].United Kingdom:Mara Con- gorithm based on random forest[J].Journal of computer ner,2017:205-237. applications,.2015,35(11)3157-3160. [7]CHEN L,MA H T,LIU S,et al.Posture estimation by [20]胡二雷，冯瑞.基于深度学习的图像检索系统).计算机 Bayesian Network with Belief Propagation[C]//TENCON 系统应用，2017,26(3)：8-19. 2013-2013 IEEE Region 10 Conference.Xi'an,China, HU Erlei,FENG Rui.Image retrieval system based on在关联关系是难以用现有的方法准确表达出来的， 通过以上方法提高跌倒检测的准确性。在分析跌倒 检测要求的基础上，针对 RNN 的特点，设计了序列 化的传感器数据、RNN 训练与检测输入表示方法， 给出了用于跌倒检测的 RNN 训练算法以及基于 RNN 的跌倒检测算法，将跌倒检测转换为输入序列 的分类问题；并在本文前期所实现的基于分布式神 经元大规模 RNN 系统的基础上，在 Spark 平台上实 现了基于 RNN 的跌倒检测系统，使用 Fall_adl_data 数据集进行了检测效果和对接近跌倒数据检测能力 的测试与分析，验证了 RNNFD 能有效提高跌倒检 测的准确率，F 值相比现有跌倒检测系统能提高 12% 和 7%，并且能够检测出接近跌倒行为。 下一步将进一步优化用于跌倒检测 RNN 的训练方法，同时考虑使用 LSTM 减少时间开销，提 高准确率。 参考文献： 吴天昊. 基于 3 轴加速度传感器及陀螺仪的老年人摔倒 识别[D]. 北京: 北京工业大学, 2013: 2–3. WU Tianhao. Identification of old people 's fall-downing based on 3-axis acceleration sensor and gyroscope[D]. Beijing, China: Beijing University of Technology, 2013: 2–3. [1] ROUGIER C, MEUNIER J, ST-ARNAUD A. 3D head tracking for fall-down detection using a single calibrated camera[J]. Image and Vision Computing, 2013, 31(3): 246–254. [2] MUBASHIR M, SHAO L, SEED L. A survey on fall-down detection: Principles and approaches[J]. Neurocomputing, 2013, 100(2): 144–152. [3] MATHIE M J, BASILAKIS J, CELLER B G. A system for monitoring posture and physical activity using acceleromet￾ers[C]//International Conference of the IEEE Engineering in Medicine and Biology Society. Istanbul, Turkey, 2001: 3654–3657. [4] 卢先领, 王洪斌, 王莹莹, 等. 一种基于加速度传感器的人 体跌倒识别方法[J]. 计算机应用研究, 2013, 30(4): 1109–1111. LU Xianling, WANG Hongbin, WANG Yingying, et al. Hu￾man fall-downing detection based on accelerometer[J]. Computer Application Research, 2013, 30(4): 1109–1111. [5] FANG-YIE LEU, CHIA-YIN KO, YI-CHEN LIN, et al. Smart Sensors Networks[M]. United Kingdom: Mara Con￾ner, 2017: 205–237. [6] CHEN L, MA H T, LIU S, et al. Posture estimation by Bayesian Network with Belief Propagation[C]//TENCON 2013-2013 IEEE Region 10 Conference. Xi’an, China, [7] 2013: 1–4. Duan K B, Keerthi S S. Which is the best multiclass SVM method? An empirical study[M]// Multiple Classifier Sys￾tems. Springer Berlin Heidelberg, 2005:278–285. [8] BREIMAN L. Random forest[J]. Machine learning, 2001, 45(1): 5–32. [9] SZ Erdogan, TT Bilgin, J Cho. Fall-down detection by us￾ing K-nearest neighbor algorithm on WSN data[C]// GLOBECOM Workshops. Houston, USA, 2011: 2054–2058. [10] WU G, XUE S. Portable preimpact fall-down detector with inertial sensors[J]. IEEE Transactions on Neural Systems & Rehabilitation Engineering, 2008, 16(2): 178–183. [11] OJETOLA O, GAURA E, BRUSEY J. Data set for fall￾down events and daily activities from inertial sensors [C]//ACM Multimedia Systems Conference. Oregon, USA, 2015: 243–248. [12] CHEN J, KWONG K, CHANG D, et al. Wearable sensors for reliable fall-down detection[C]//International Confer￾ence of the IEEE, NEW YORK, USA, 2006: 3551–3554. [13] LI Q, STANKOVIC J A, HANSON M A, et al. Accurate fast fall-down detection using gyroscopes and acceleromet￾er-derived posture information[C]//Sixth International Workshop on Wearable and Implantable Body Sensor Net￾works. CA, USA, 2009: 138–143. [14] VAIDEHI V, GANAPATHY K, MOHAN K, et al. Video based automatic fall-down detection in indoor environ￾ment[C]// International Conference on Recent Trends in In￾formation Technology. Chennai, INDIA, 2011: 1016– 1020. [15] BOSCH-JORGE M, SÁNCHEZ-SALMERÓN A J, ÁNGEL VALERA, et al. Fall-down detection based on the gravity vector using a wide-angle camera[J]. Expert Sys￾tems with Applications, 2014, 41(17): 7980–7986. [16] 佟丽娜, 宋全军, 葛运建. 基于时序分析的人体摔倒预测 方法[J]. 模式识别与人工智能, 2012, 25(2): 273–279. TONG Linna, SONG Quanjun, GE Yunjian. Time series analysis based human fall-down prediction method[J]. Pat￾tern Recognition & Artificial Intelligence, 2012, 25(2): 273–279. [17] GIBSON R M, AMIRA A, CASASECA-DE-LA￾HIGUERA P, et al. An efficient user-customisable mul￾tiresolution classifier fall-down detection and diagnostic system[C]//International Conference on Microelectronics. Changchun, China, 2015: 228–231. [18] LUO D, LUO H, SCHOOL I. Fall-down detection al￾gorithm based on random forest[J]. Journal of computer applications, 2015, 35(11): 3157–3160. [19] 胡二雷, 冯瑞. 基于深度学习的图像检索系统[J]. 计算机 系统应用, 2017, 26(3): 8–19. HU Erlei, FENG Rui. Image retrieval system based on [20] ·386· 智 能 系 统 学 报 第 13 卷