TABLE 6.
Fall detection and human activity recognition.
| Reference | AI Algorithm Best Achieved accuracy | Data Acquisition | Task |
|---|---|---|---|
| Chang et al. (2021) | HMM with OpenPose | Two cameras | Fall risk assessment. Evaluation of imbalanced gait |
| Shioiri et al. (2021) | SVM, 79% | micro-Doppler radar | Classification of gait differences associated with fall risk |
| CNN 73% | |||
| Lu et al. (2021b) | SOT, improved accuracy by 6% | Public HAR datasets UCI-DSADS, UCI-HAR, USC–HAD, PAMAP2 | Cross-domain HAR, utilizing transfer learning from auxiliary labeled data |
| Mori et al. (2020) | NN | 11 men, TW, induced disturbances | Predict falls caused by an unexpected disturbance in time for CD to deploy |
| Chelli and Pätzold, (2019) | ANN, KNN, QSVM, EBT. fall detection = 100%, false alarms = 0, ARA = 97.7% | Wearable sensors Public datasets (Anguita et al., 2013) and (Ojetola et al., 2015) that record falls, near-falls, and 7 ADL | ADL recognition. Fall detection |
| Kondragunta et al. (2019) | OpenPose for 2D pose estimation | Kinect images and sensor gait data from 250 subjects, 4 times, over 3 years | Estimation of Gait Parameters for Elderly Care from 3D Pose |
| Weiss et al. (2019) | RF, DT, KNN with K = 5. EER = 9.3 by RF. RF performs best in most of the sensor combinations | 51 subjects, 18 ADL. Smartphones in right pocket and smartwatch on the dominant hand | Continuous biometrics authentication and identification on smartphones or smartwatches. |
| Santoyo-Ramón et al. (2018) | SVM, KNN, NB, DT. Error 14.162% by SVM. | Inertial sensors. 19 subjects at home, 3 falls and 11 ADL | Wearable Fall Detection System |
| Yang et al. (2018) | CSVD-NMF. 96.8% occupancy detection. 90.6% activity recognition | WiFi-enabled CSI measurements of 5 ADL | Device-Free Occupancy Sensing and activity recognition |
| Yu et al. (2018) | Gaussian HMM. Sensitivity of 0.992. Positive predictive value of 0.981 | Own data. 200 fall events and 385 normal activities | Fall detection system |
| Seyfioğlu et al. (2018) | DCAE vs. CNN, SVM, AE. | micro-Doppler signatures | Radar-based activity recognition |
| Xi et al. (2017) | ARA = 97.35% by GK-SVM. FD: sensitivity 98.70% and specificity 98.59% by GK-FDA. | 3 subjects, 7 ADL Wireless wearable sEMG sensors | Automatic activity recognition and fall detection |
| Daher et al. (2017) | HCM-SFS on fused GRF and accelerometer data. ARA> 90% on all 5 ADL. | Force sensors and accelerometers under intelligent tiles. 6 subjects, 5 ADL | Fall detection and ADL recognition in independent living senior apartments |
| Hakim et al. (2017) | SVM, NN, DT, DA. 99% by SVM. | Smart phone IMU. 8 healthy subjects, 4 fall events, 6 ADL | ADL recognition and threshold-based fall detection |
| Gao et al. (2017) | SVM | WiFi CSI measurements | Device-free wireless localization and activity recognition |
| Wu et al. (2015a) | Sparse BC+RVM. | 2 falling, 6 ADL, Spectrograms from continuous-wave radar | Radar-based Fall Detection |
| (Wannenburg and Malekian 2017) | KNN, kStar, HMM, SVM, DTC, RF, NB LR, ANN | smartphone | Activity recognition |
| Ngo et al. (2015) | SVM, KNN | inertial sensor | Recognition for similar gait action classes |
| Semwal et al. (2015) | k-means and KNN ANN + PCA | vision and sensor-based gait data | Abnormal gait detection |
| Ma et al. (2014b) | Variable-length PSO+ELM. 91.15% sensitivity, 77.14% specificity, and 86.83% accuracy | 10 young subjects, intentionally falling, and 6 ADL Kinect depth camera | Shape-based fall detection that is invariant to human translation, rotation, scaling and action length |
| Özdemir and Barshan, (2014) | KNN, LSM Over 99% | 14 subjects, 20 falls, 16 ADL, 6 wearable sensors | Automated fall detection system |
| (Mannini and Sabatini, 2012) | HMM | wireless IMU and an optical motion analysis system | Gait phase detection and walking/jogging discrimination |
Legend: Quadratic SVM (QSVM), HCM (Histogram Comparison Method), Sequential Forward Selection (SFS), Least squares method (LSM), Gaussian Kernel Fisher Discriminant Analysis (GK-FDA), Non-Negative Matrix Factorization (NMF), Class Estimated Basis Space Singular Value Decomposition (CSVD), Equal Error Rate (EER), Relevance Vector Machine (RVM), Gaussian Kernel SVM (GK-SVM), Substructural Optimal Transport (SOT), Channel State Information (CSI).
Datasets: UCI-DSADS (Anguita et al., 2012) UCI-HAR (Barshan and Yüksek, 2014), USC–HAD (Zhang and Sawchuk, 2012), PAMAP2 (Reiss and Stricker, 2012).