Table 1 A comparison of modeling approaches in HAR for physiotherapy.
From: Spike train analysis in rehabilitation movement classification using deep learning approach
Approach | Input representation | Strengths | Limitations | References |
|---|---|---|---|---|
CNN (image/video) | RGB frames, depth maps | Strong in spatial feature extraction, widely benchmarked | High dimensionality, sensitive to lighting/occlusion | |
Skeleton-based GCN | Skeleton joints as graph nodes | Captures joint dependencies; spatial–temporal modeling | Requires accurate joint detection; less robust to noise | |
ST-GCN | Skeleton sequences with temporal graph edges | Models spatial and temporal dynamics simultaneously | Computationally heavy; sensitive to missing joints | |
RNN/LSTM | Skeleton joint sequences | Effective for temporal dependencies | Struggles with long-range dependencies | |
Time–frequency/Heatmaps | Transformed motion signals | Rich spectral–temporal features; interpretable | Preprocessing overhead | |
Spike train encoding (Proposed) | Rasterized spike trains from skeletal data | Bio-inspired, temporally sparse, highlights movement correctness | Novel approach, requires adaptation of classifiers | Proposed |
