Table 1.
Summary of related works based on conventional machine learning methods.
Table 2.
Summary of related works based on deep learning methods.
Fig 1.
Proposed framework: (a) Training of an SVM classifier using ResNet-50 features, (b) Posture classification using the trained SVM model and ResNet-50 features.
Fig 2.
A residual block.
Fig 3.
Key frames of MCF dataset showing different postures: (a) Standing or walking, (b) Lying on ground, (c) Moving down or sitting, (d) Standing up.
Table 3.
Number of frames for each posture of four datasets namely MCF, URFD, UPFD-Four Postures (UPFD-4P), and UPFD-Two Postures (UPFD-2P).
Fig 4.
Key frames of UR Fall dataset with varying postures: (a) standing or walking, (b) lying on ground, (c) moving down or falling, (d) standing up.
Fig 5.
Key frames of UPFD dataset with varying postures: (a) standing or walking, (b) lying on the ground, (c) moving down or falling, (d) standing up or jumping.
Table 4.
Comparison of the proposed approach (PA) in terms of accuracy (%) with existing architectures on the MCF dataset.
Table 5.
Comparison of proposed approach (PA) in terms of time (sec) with existing architectures on MCF dataset.
Table 6.
Comparison of existing approaches with proposed approach (PA) in terms of accuracy (%) on MCF dataset.
Fig 6.
Comparison of the proposed approach (PA) in terms of accuracy (%) with existing architectures on URFD dataset and UP-Fall detection dataset.
Fig 7.
Comparison of proposed approach (PA) in terms of time (sec) with existing architectures on URFD dataset and UP-Fall detection dataset.
Table 7.
Comparison of existing approaches with proposed approach (PA) in terms of % accuracy on URFD dataset.
Table 8.
Comparison of existing approaches with proposed approach (PA) in terms of accuracy (%) on UPFD dataset.
Table 9.
Parameters used for comparison of Decision Tree, Random Forest, KNN, AdaBoost, MLP, and SVM.
Table 10.
Comparison of variants of the proposed approach (PA) in terms of accuracy (%) on four datasets namely MCF (Camera 1 to Camera 8), URFD, UPFD (Two Postures), and UPFD (Four Postures).
Fig 8.
Features (first two principal components (A1 & A2)) of pooling layer of proposed approach of URFD dataset for four human poses including standing or walking, lying on the ground, moving down or falling, and standing up using (a) Scatter plot, (b) Linear projection.
Table 11.
Within-class (average distance) and between-class(average linkage) distances for URFD dataset (ResNet-50 features).
Fig 9.
VGG-19 low-level feature visualization of URFD dataset for human pose standing or walking.
Fig 10.
ResNet-50 low-level feature visualization of URFD dataset for human pose standing or walking.